Strands Decider

Exact model and requirements

Artifact:

Strands Agents; Qwen base model / Strands Decider 2B hobson v19 base + LoRA + pointer head · checkpoint StrandsAgents/strands-decider-2B-hobson-v19

Publisher revision:

bb282d786bc251fd4e3068de3ada9ddbb38127cd

Artifact digest:

Composite checkpoint and separately pinned base have multiple hashes; exact component SHA256 values are pinned in auxiliary artifacts and recipe.

Download size:

4.3 GiB

Memory requirement (unknown):

Unknown — No exact full-process CPU peak for pinned base/adapter/head workload.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no model sharding. CPU-only torch execution; no accelerator or disk offload configured.

KV cache:

Torch runtime manages attention and recurrent state. Prefix cache disabled; exact state/cache precision and peak bytes not separately measured.

Workload:

4,096 token context; Unknown; Not applicable; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Exact v19 exported tokenizer · bb282d786bc251fd4e3068de3ada9ddbb38127cd

Sampling / reasoning:

No autoregressive generation sampler. v19 per-kind calibration: noul 0.9107136998460428; choice 0.734189596436441; score 1.32780942142348. No text reasoning trace or decode loop. Typed classification/readout over a state and explicit question options.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Strands v19 and its runtime declare Apache License 2.0; the pinned Qwen base is also Apache 2.0. Read and retain applicable license, copyright, attribution and modification notices when distributing covered work. Dataset attribution in the v19 license does not make unrelated dataset uses automatically permitted.

• Qwen base weight shard · 4.24 GiB · revision b1485b2fa6dfa1287294f269f5fb618e03d52d7c
• Qwen base config · 0 GiB · revision b1485b2fa6dfa1287294f269f5fb618e03d52d7c
• lora/adapter_model.safetensors · 0.06 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• lora/adapter_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• head.safetensors · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• hobson_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• provenance.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• tokenizer.json · 0.02 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• tokenizer_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd

• Base download includes the original multimodal checkpoint shard, while text loader keeps only its decoder. Download bytes do not establish runtime working memory.
• FP32 CPU torso plus head, adapter, tokenizer, recurrent/attention state, packages and OS require more than just on-disk weight size. No generic token throughput claim is applicable.

Install the runtime

1. 1. Install the exact CPU source runtime

   Create a dedicated virtual environment and install the pinned source tree. Its pyproject declares torch >=2.7, transformers >=5.15,<6, PEFT >=0.21 and related dependencies. No CUDA-specific extra is selected. Dependency versions are not a fully locked environment; retain the resolved freeze with any execution evidence.

   CommandCopy command

   sudo apt-get update
   sudo apt-get install -y python3 python3-venv python3-pip git curl ca-certificates
   mkdir -p "$HOME/strands-v19-hosting"
   cd "$HOME/strands-v19-hosting"
   python3 -m venv .venv
   source .venv/bin/activate
   python -m pip install --upgrade pip
   git clone https://github.com/strands-labs/strands-decider.git runtime
   git -C runtime checkout --detach 6039cf8b64c6074e75804000c07ac3d4cce795f7
   python -m pip install -e './runtime[cpu]'
   python -m pip freeze > installed-packages.txt

   Expected result: Source checkout is exactly pinned; installation succeeds and installed-packages.txt records resolved dependencies. No inference has run yet.

Download the model

1. 1. Freeze the complete v19 export and its documented base revision

   Download the full v19 snapshot into a local folder so the server never resolves checkpoint main. Prefetch the published base revision into the normal Hub cache. The base pin is explicitly inferred in provenance, not proof of the exact historical training revision. Keeping the full snapshot allows the publisher manifest verification.

   CommandCopy command

   cd "$HOME/strands-v19-hosting"
   source .venv/bin/activate
   python - <<'PYDOWNLOAD'
   from huggingface_hub import snapshot_download
   snapshot_download(repo_id='StrandsAgents/strands-decider-2B-hobson-v19', revision='bb282d786bc251fd4e3068de3ada9ddbb38127cd', local_dir='./checkpoint')
   snapshot_download(repo_id='Qwen/Qwen3.5-2B-Base', revision='b1485b2fa6dfa1287294f269f5fb618e03d52d7c')
   PYDOWNLOAD
   python -m strands_decider.hf_export verify ./checkpoint

   Expected result: Complete checkpoint files match MANIFEST.sha256. Base cache resolves the explicit revision. Stop on missing adapter/head/config/tokenizer or verification failure.

2. 2. Verify critical adapter, head and base weight bytes

   The checkpoint manifest and explicit hashes bind the v19 adapter/head; verify the single base shard in the pinned Hub cache as well. No file is modified to substitute a different base.

   CommandCopy command

   cd "$HOME/strands-v19-hosting"
   sha256sum --check <<'SHACHECK'
   701bdb895887097f7954ec7eb06f7937d3b790035b5462195eb26abd80a4aebc  checkpoint/lora/adapter_model.safetensors
   daad0727152b6185447cee36f78230c144312feb7b19f02749b19645238b5287  checkpoint/head.safetensors
   SHACHECK
   source .venv/bin/activate
   python - <<'PYBASE'
   from huggingface_hub import hf_hub_download
   import hashlib
   f=hf_hub_download(repo_id='Qwen/Qwen3.5-2B-Base',filename='model.safetensors-00001-of-00001.safetensors',revision='b1485b2fa6dfa1287294f269f5fb618e03d52d7c')
   h=hashlib.sha256()
   with open(f,'rb') as stream:
       for chunk in iter(lambda: stream.read(8*1024*1024), b''):
           h.update(chunk)
   assert h.hexdigest() == '928acbf11878c32185bbd863514d191769285065ab9ea14fbfe431303f5fdf2d', 'Base weight mismatch'
   print('Pinned base weight SHA256 OK')
   PYBASE

   Expected result: Adapter and head report OK; pinned base weight SHA256 OK. This verifies bytes, not runtime fit or accuracy.

Configure

1. 1. Check base resolution and CPU settings

   The pinned loader reads hobson_config.json for this legacy export. It takes config.base_revision first, otherwise matching provenance.base_model_revision, then passes that revision to AutoConfig and Qwen3_5ForCausalLM. It attaches PEFT from lora/ and restores head.safetensors. Confirm those identities before starting. OMP/MKL four-thread values are a runtime allocation; use one client request at a time for this walkthrough.

   CommandCopy command

   cd "$HOME/strands-v19-hosting"
   source .venv/bin/activate
   python - <<'PYCHECK'
   import json
   from strands_decider.modeling import StrandsDeciderConfig,base_revision,config_path
   p='./checkpoint'
   c=StrandsDeciderConfig.from_json(config_path(p))
   assert c.base_model == 'Qwen/Qwen3.5-2B-Base'
   assert base_revision(p,c) == 'b1485b2fa6dfa1287294f269f5fb618e03d52d7c'
   assert c.max_length == 4096 and c.num_slots == 24 and c.use_lora
   print('v19 config and published inferred base revision match')
   PYCHECK

   Expected result: Config window 4096, slots 24 and published inferred base revision match. This check does not load model weights into inference.

Start the model

1. 1. Serve typed text decisions on CPU

   Run in a foreground terminal. --strict-window rejects overlong input instead of truncating its end. --max-batch 1 processes one question per forward; three questions still receive three typed answers. Prefix caching is explicitly disabled for this basic configuration. This setting does not enforce a global one-client limit; keep the walkthrough at one request at a time. No text-generation sampler or API credential is used.

   CommandCopy command

   cd "$HOME/strands-v19-hosting"
   source .venv/bin/activate
   export OMP_NUM_THREADS=4 MKL_NUM_THREADS=4
   strands-decider serve ./checkpoint --host 127.0.0.1 --port 8000 --device cpu --model-name strands-v19 --strict-window --max-batch 1 --no-prefix-cache

   Expected result: Weights, adapter and head load; health becomes available on loopback. Stop if files, base identity or loading fail.

Verify the setup

1. 1. Verify health and all three decision primitives

   Use a second terminal on the server. No API key is required because the native server has no authentication. Do not expose it publicly. Health must identify this v19 setup, CPU, max_length 4096, vision false and prefix_cache false. The example checks an obvious billing triage case; probabilities/confidence are not guaranteed to equal the old model-card example.

   CommandCopy command

   cd "$HOME/strands-v19-hosting"
   curl --fail http://127.0.0.1:8000/health
   cat > decision-request.json <<'REQUEST'
   {
     "state": "Help! My payouts have been failing for 3 days!",
     "questions": {
       "urgent": {
         "type": "noul",
         "instructions": "Does this convey urgency?"
       },
       "team": {
         "type": "choice",
         "instructions": "Which team should handle this?",
         "criteria": {
           "billing": "payout and payment problems",
           "sales": "buying a product",
           "retail": "physical store requests"
         }
       },
       "frustration": {
         "type": "score",
         "instructions": "How frustrated is the writer?",
         "criteria": [
           "calm",
           "frustrated",
           "very frustrated"
         ]
       }
     }
   }
   REQUEST
   curl --fail http://127.0.0.1:8000/v1/systemone -H 'Content-Type: application/json' --data-binary @decision-request.json --output decision-response.json
   cat decision-response.json

   Expected result: HTTP 200 response with model strands-v19 and answers.urgent.noul in [0,1], answers.team.choice naming a declared team with probabilities/confidence, and answers.frustration.score in [0,2] with level legend/probabilities/confidence. The obvious team should be billing; inspect it. No choices[0].message.content or generated-token claim is expected. Wrong labels require task validation, not blind automation.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• This is a typed decision model, not a general text-chat or token-generation endpoint. noul gives P(true); choice gives a named option; score gives an expectation over ordered level indices. Generated output tokens are not applicable.
• The frozen v19 export is a LoRA adapter plus pointer head over Qwen3.5-2B-Base. All three are required. The source loader preserves v19’s legacy config and reads the published inferred base pin; training hosts did not pin that base, so this is a reproducible published inference choice, not certain historical training provenance.
• Checkpoint max_length is 4096. Its metadata num_slots is 24, but v19 uses a pointer head, so num_slots does not impose a choice-option ceiling. The request schema permits 2–255 choice options, still subject to the rendered context window. Score rubrics use 2–10 ascending levels. Strict-window overflow returns HTTP 422; verify the intended labels and token budget.
• CPU execution upcasts its half-precision torso to FP32 and uses CPU-specific kernels; the head is FP32. No full process memory, per-request latency, throughput or cost was measured. Do not infer peak RAM from download bytes or transfer another version/device benchmark.
• Per-kind calibration temperatures are from v19 config (noul 0.9107136998460428, choice 0.734189596436441, score 1.32780942142348), not generation sampling settings. Calibration was fitted to held-out short tasks; measure your own traffic before trusting a confidence threshold.
• Publisher warns question wording is read less than state/options, long multi-step documents are weak, and score/noul transfer poorly to unfamiliar rubrics. Do not equate a high returned confidence with safety or correctness. Validate the intended task and failure handling.
• This item is text-only. Requests carrying images are rejected when the server is not started with --vision. Supported v19 visual typed-decision jobs are a separate required guide scope; this item does not claim to complete them.
• The current pinned source is newer than the original v19 card. No newer v1 benchmark/calibration results or exact old example probabilities are transferred. Dependency ranges are declared but not fully locked; preserve installed-packages.txt with actual run evidence.

Missing adapter/head, bad checksum or unexpected base model

Stop. Restore the complete exact v19 snapshot and manifest, including lora/, head, config, tokenizer and provenance. Verify critical SHA256 values and the inferred base revision. Do not serve just the adapter or silently fall back to main.

Import error or unknown Qwen3.5 class

Confirm pinned source checkout, active virtual environment and declared dependency constraints. Avoid assuming PyPI 0.1.0 includes current device extras. Keep the resolver log and pip freeze; do not substitute a generative chat runtime.

HTTP 422 or slow/out-of-memory request

Check the request schema: 2–255 named choice options and 2–10 ordered score levels. v19’s pointer head is not capped by metadata num_slots 24. Keep rendered state/questions/options inside the strict 4096 window and process one request at a time. Free/provision resources or shorten input; record changes. No measured fit or speed is promised.

High-confidence wrong choice or unsuitable score

Check that state/options express the intended decision, simplify long multi-step documents and validate on representative labeled traffic. Adjust task design or train/calibrate separately; do not enable automatic action based solely on confidence.

Connection refused

Check foreground server logs, CPU health and loopback port 8000. Run the client on the same trusted desktop.

Stop the model

1. 1. Stop the server and release the model

   Press Ctrl+C in the foreground server terminal. Confirm localhost:8000 no longer serves decisions.

   Expected result: Server stops; loopback endpoint no longer responds.

Uninstall

1. 1. Remove only dedicated walkthrough files

   Back up wanted request/results and installed-packages.txt. Remove the dedicated folder after confirming its path. The shared Hugging Face base cache remains; list/manage it separately with the installed Hub CLI and remove only unwanted exact repositories/revisions, as other workloads may share it.

   CommandCopy command

   rm -r -- "$HOME/strands-v19-hosting"

   Expected result: Dedicated runtime/venv/checkpoint files removed. Shared Hub cache and system packages are not blindly deleted.

Related guides

• https://opentools.ai/llms/strands-decider

Exact model and requirements

Artifact:

Strands Agents; Qwen base model / Strands Decider 2B hobson v19 base + LoRA + pointer head · checkpoint StrandsAgents/strands-decider-2B-hobson-v19

Publisher revision:

bb282d786bc251fd4e3068de3ada9ddbb38127cd

Artifact digest:

Composite checkpoint and separately pinned base have multiple hashes; exact component SHA256 values are pinned in auxiliary artifacts and recipe.

Download size:

4.3 GiB

Memory requirement (unknown):

Unknown — No exact full-process CPU peak for pinned base/adapter/head workload.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no model sharding. CPU-only torch execution; no accelerator or disk offload configured.

KV cache:

Torch runtime manages attention and recurrent state. Prefix cache disabled; exact state/cache precision and peak bytes not separately measured.

Workload:

4,096 token context; Unknown; Not applicable; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Exact v19 exported tokenizer · bb282d786bc251fd4e3068de3ada9ddbb38127cd

Sampling / reasoning:

No autoregressive generation sampler. v19 per-kind calibration: noul 0.9107136998460428; choice 0.734189596436441; score 1.32780942142348. No text reasoning trace or decode loop. Typed classification/readout over a state and explicit question options.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Strands v19 and its runtime declare Apache License 2.0; the pinned Qwen base is also Apache 2.0. Read and retain applicable license, copyright, attribution and modification notices when distributing covered work. Dataset attribution in the v19 license does not make unrelated dataset uses automatically permitted.

• Qwen base weight shard · 4.24 GiB · revision b1485b2fa6dfa1287294f269f5fb618e03d52d7c
• Qwen base config · 0 GiB · revision b1485b2fa6dfa1287294f269f5fb618e03d52d7c
• lora/adapter_model.safetensors · 0.06 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• lora/adapter_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• head.safetensors · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• hobson_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• provenance.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• tokenizer.json · 0.02 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• tokenizer_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd

• Base download includes the original multimodal checkpoint shard, while text loader keeps only its decoder. Download bytes do not establish runtime working memory.
• FP32 CPU torso plus head, adapter, tokenizer, recurrent/attention state, packages and OS require more than just on-disk weight size. No generic token throughput claim is applicable.

Install the runtime

1. 1. Install the exact CPU source runtime

   Create a dedicated virtual environment and install the pinned source tree. Its pyproject declares torch >=2.7, transformers >=5.15,<6, PEFT >=0.21 and related dependencies. No CUDA-specific extra is selected. Dependency versions are not a fully locked environment; retain the resolved freeze with any execution evidence.

   CommandCopy command

   sudo apt-get update
   sudo apt-get install -y python3 python3-venv python3-pip git curl ca-certificates
   mkdir -p "$HOME/strands-v19-hosting"
   cd "$HOME/strands-v19-hosting"
   python3 -m venv .venv
   source .venv/bin/activate
   python -m pip install --upgrade pip
   git clone https://github.com/strands-labs/strands-decider.git runtime
   git -C runtime checkout --detach 6039cf8b64c6074e75804000c07ac3d4cce795f7
   python -m pip install -e './runtime[cpu]'
   python -m pip freeze > installed-packages.txt

   Expected result: Source checkout is exactly pinned; installation succeeds and installed-packages.txt records resolved dependencies. No inference has run yet.

Download the model

1. 1. Freeze the complete v19 export and its documented base revision

   Download the full v19 snapshot into a local folder so the server never resolves checkpoint main. Prefetch the published base revision into the normal Hub cache. The base pin is explicitly inferred in provenance, not proof of the exact historical training revision. Keeping the full snapshot allows the publisher manifest verification.

   CommandCopy command

   cd "$HOME/strands-v19-hosting"
   source .venv/bin/activate
   python - <<'PYDOWNLOAD'
   from huggingface_hub import snapshot_download
   snapshot_download(repo_id='StrandsAgents/strands-decider-2B-hobson-v19', revision='bb282d786bc251fd4e3068de3ada9ddbb38127cd', local_dir='./checkpoint')
   snapshot_download(repo_id='Qwen/Qwen3.5-2B-Base', revision='b1485b2fa6dfa1287294f269f5fb618e03d52d7c')
   PYDOWNLOAD
   python -m strands_decider.hf_export verify ./checkpoint

   Expected result: Complete checkpoint files match MANIFEST.sha256. Base cache resolves the explicit revision. Stop on missing adapter/head/config/tokenizer or verification failure.

2. 2. Verify critical adapter, head and base weight bytes

   The checkpoint manifest and explicit hashes bind the v19 adapter/head; verify the single base shard in the pinned Hub cache as well. No file is modified to substitute a different base.

   CommandCopy command

   cd "$HOME/strands-v19-hosting"
   sha256sum --check <<'SHACHECK'
   701bdb895887097f7954ec7eb06f7937d3b790035b5462195eb26abd80a4aebc  checkpoint/lora/adapter_model.safetensors
   daad0727152b6185447cee36f78230c144312feb7b19f02749b19645238b5287  checkpoint/head.safetensors
   SHACHECK
   source .venv/bin/activate
   python - <<'PYBASE'
   from huggingface_hub import hf_hub_download
   import hashlib
   f=hf_hub_download(repo_id='Qwen/Qwen3.5-2B-Base',filename='model.safetensors-00001-of-00001.safetensors',revision='b1485b2fa6dfa1287294f269f5fb618e03d52d7c')
   h=hashlib.sha256()
   with open(f,'rb') as stream:
       for chunk in iter(lambda: stream.read(8*1024*1024), b''):
           h.update(chunk)
   assert h.hexdigest() == '928acbf11878c32185bbd863514d191769285065ab9ea14fbfe431303f5fdf2d', 'Base weight mismatch'
   print('Pinned base weight SHA256 OK')
   PYBASE

   Expected result: Adapter and head report OK; pinned base weight SHA256 OK. This verifies bytes, not runtime fit or accuracy.

Configure

1. 1. Check base resolution and CPU settings

   The pinned loader reads hobson_config.json for this legacy export. It takes config.base_revision first, otherwise matching provenance.base_model_revision, then passes that revision to AutoConfig and Qwen3_5ForCausalLM. It attaches PEFT from lora/ and restores head.safetensors. Confirm those identities before starting. OMP/MKL four-thread values are a runtime allocation; use one client request at a time for this walkthrough.

   CommandCopy command

   cd "$HOME/strands-v19-hosting"
   source .venv/bin/activate
   python - <<'PYCHECK'
   import json
   from strands_decider.modeling import StrandsDeciderConfig,base_revision,config_path
   p='./checkpoint'
   c=StrandsDeciderConfig.from_json(config_path(p))
   assert c.base_model == 'Qwen/Qwen3.5-2B-Base'
   assert base_revision(p,c) == 'b1485b2fa6dfa1287294f269f5fb618e03d52d7c'
   assert c.max_length == 4096 and c.num_slots == 24 and c.use_lora
   print('v19 config and published inferred base revision match')
   PYCHECK

   Expected result: Config window 4096, slots 24 and published inferred base revision match. This check does not load model weights into inference.

Start the model

1. 1. Serve typed text decisions on CPU

   Run in a foreground terminal. --strict-window rejects overlong input instead of truncating its end. --max-batch 1 processes one question per forward; three questions still receive three typed answers. Prefix caching is explicitly disabled for this basic configuration. This setting does not enforce a global one-client limit; keep the walkthrough at one request at a time. No text-generation sampler or API credential is used.

   CommandCopy command

   cd "$HOME/strands-v19-hosting"
   source .venv/bin/activate
   export OMP_NUM_THREADS=4 MKL_NUM_THREADS=4
   strands-decider serve ./checkpoint --host 127.0.0.1 --port 8000 --device cpu --model-name strands-v19 --strict-window --max-batch 1 --no-prefix-cache

   Expected result: Weights, adapter and head load; health becomes available on loopback. Stop if files, base identity or loading fail.

Verify the setup

1. 1. Verify health and all three decision primitives

   Use a second terminal on the server. No API key is required because the native server has no authentication. Do not expose it publicly. Health must identify this v19 setup, CPU, max_length 4096, vision false and prefix_cache false. The example checks an obvious billing triage case; probabilities/confidence are not guaranteed to equal the old model-card example.

   CommandCopy command

   cd "$HOME/strands-v19-hosting"
   curl --fail http://127.0.0.1:8000/health
   cat > decision-request.json <<'REQUEST'
   {
     "state": "Help! My payouts have been failing for 3 days!",
     "questions": {
       "urgent": {
         "type": "noul",
         "instructions": "Does this convey urgency?"
       },
       "team": {
         "type": "choice",
         "instructions": "Which team should handle this?",
         "criteria": {
           "billing": "payout and payment problems",
           "sales": "buying a product",
           "retail": "physical store requests"
         }
       },
       "frustration": {
         "type": "score",
         "instructions": "How frustrated is the writer?",
         "criteria": [
           "calm",
           "frustrated",
           "very frustrated"
         ]
       }
     }
   }
   REQUEST
   curl --fail http://127.0.0.1:8000/v1/systemone -H 'Content-Type: application/json' --data-binary @decision-request.json --output decision-response.json
   cat decision-response.json

   Expected result: HTTP 200 response with model strands-v19 and answers.urgent.noul in [0,1], answers.team.choice naming a declared team with probabilities/confidence, and answers.frustration.score in [0,2] with level legend/probabilities/confidence. The obvious team should be billing; inspect it. No choices[0].message.content or generated-token claim is expected. Wrong labels require task validation, not blind automation.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• This is a typed decision model, not a general text-chat or token-generation endpoint. noul gives P(true); choice gives a named option; score gives an expectation over ordered level indices. Generated output tokens are not applicable.
• The frozen v19 export is a LoRA adapter plus pointer head over Qwen3.5-2B-Base. All three are required. The source loader preserves v19’s legacy config and reads the published inferred base pin; training hosts did not pin that base, so this is a reproducible published inference choice, not certain historical training provenance.
• Checkpoint max_length is 4096. Its metadata num_slots is 24, but v19 uses a pointer head, so num_slots does not impose a choice-option ceiling. The request schema permits 2–255 choice options, still subject to the rendered context window. Score rubrics use 2–10 ascending levels. Strict-window overflow returns HTTP 422; verify the intended labels and token budget.
• CPU execution upcasts its half-precision torso to FP32 and uses CPU-specific kernels; the head is FP32. No full process memory, per-request latency, throughput or cost was measured. Do not infer peak RAM from download bytes or transfer another version/device benchmark.
• Per-kind calibration temperatures are from v19 config (noul 0.9107136998460428, choice 0.734189596436441, score 1.32780942142348), not generation sampling settings. Calibration was fitted to held-out short tasks; measure your own traffic before trusting a confidence threshold.
• Publisher warns question wording is read less than state/options, long multi-step documents are weak, and score/noul transfer poorly to unfamiliar rubrics. Do not equate a high returned confidence with safety or correctness. Validate the intended task and failure handling.
• This item is text-only. Requests carrying images are rejected when the server is not started with --vision. Supported v19 visual typed-decision jobs are a separate required guide scope; this item does not claim to complete them.
• The current pinned source is newer than the original v19 card. No newer v1 benchmark/calibration results or exact old example probabilities are transferred. Dependency ranges are declared but not fully locked; preserve installed-packages.txt with actual run evidence.

Missing adapter/head, bad checksum or unexpected base model

Stop. Restore the complete exact v19 snapshot and manifest, including lora/, head, config, tokenizer and provenance. Verify critical SHA256 values and the inferred base revision. Do not serve just the adapter or silently fall back to main.

Import error or unknown Qwen3.5 class

Confirm pinned source checkout, active virtual environment and declared dependency constraints. Avoid assuming PyPI 0.1.0 includes current device extras. Keep the resolver log and pip freeze; do not substitute a generative chat runtime.

HTTP 422 or slow/out-of-memory request

Check the request schema: 2–255 named choice options and 2–10 ordered score levels. v19’s pointer head is not capped by metadata num_slots 24. Keep rendered state/questions/options inside the strict 4096 window and process one request at a time. Free/provision resources or shorten input; record changes. No measured fit or speed is promised.

High-confidence wrong choice or unsuitable score

Check that state/options express the intended decision, simplify long multi-step documents and validate on representative labeled traffic. Adjust task design or train/calibrate separately; do not enable automatic action based solely on confidence.

Connection refused

Check foreground server logs, CPU health and loopback port 8000. Verify the SSH tunnel to the intended VM; do not expose the unauthenticated port.

Stop the model

1. 1. Stop the server and release the model

   Press Ctrl+C in the foreground server terminal. Close the SSH tunnel with Ctrl+C; stopping only this process does not stop EC2 billing.

   Expected result: Server stops; loopback endpoint no longer responds.

Uninstall

1. 1. Remove only dedicated walkthrough files

   Back up wanted request/results and installed-packages.txt. Remove the dedicated folder after confirming its path. The shared Hugging Face base cache remains; list/manage it separately with the installed Hub CLI and remove only unwanted exact repositories/revisions, as other workloads may share it.

   CommandCopy command

   rm -r -- "$HOME/strands-v19-hosting"

   Expected result: Dedicated runtime/venv/checkpoint files removed. Shared Hub cache and system packages are not blindly deleted.

Shut down cloud resources

1. 1. End cloud billing resources

   In EC2 us-east-1, stop for later reuse or terminate this exact VM. Review Delete on termination for its root volume. Delete unwanted detached EBS volumes, snapshots and reserved/Elastic/public IP resources. Stopped VM storage and retained resources may still bill. Check Bills/Cost Explorer later for final usage.

   Expected result: Exact instance is stopped/terminated as intended; no unwanted attached/retained billed resources remain. Instance compute stops only after its lifecycle transition, not when llama-server stops.

Related guides

• https://opentools.ai/llms/strands-decider

Exact model and requirements

Artifact:

Strands Agents; Qwen base model / Strands Decider 2B hobson v19 base + LoRA + pointer head · checkpoint StrandsAgents/strands-decider-2B-hobson-v19

Publisher revision:

bb282d786bc251fd4e3068de3ada9ddbb38127cd

Artifact digest:

Composite checkpoint and separately pinned base have multiple hashes; exact component SHA256 values are pinned in auxiliary artifacts and recipe.

Download size:

4.3 GiB

Memory requirement (unknown):

Unknown — No exact full-process CPU peak for pinned base/adapter/head workload.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no model sharding. CPU-only torch execution; no accelerator or disk offload configured.

KV cache:

Torch runtime manages attention cache and recurrent state. Native vision path uses a prefix/suffix cache; exact state/cache precision and peak bytes are not separately measured.

Workload:

4,096 token context; Unknown; Not applicable; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Exact v19 exported tokenizer · bb282d786bc251fd4e3068de3ada9ddbb38127cd

Sampling / reasoning:

No generative sampler. Frozen per-kind calibration temperatures from v19 config; typed noul/choice/score readouts. No generated thinking or prose. Native vision prefix/suffix path uses a cache unconditionally; --max-batch1 re-encodes the image for each question chunk. Request contains image plus typed questions.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Strands v19 and its runtime declare Apache License 2.0; the pinned Qwen base is also Apache 2.0. Read and retain applicable license, copyright, attribution and modification notices when distributing covered work. Dataset attribution in the v19 license does not make unrelated dataset uses automatically permitted.

• Qwen base weight shard · 4.24 GiB · revision b1485b2fa6dfa1287294f269f5fb618e03d52d7c
• Qwen base config · 0 GiB · revision b1485b2fa6dfa1287294f269f5fb618e03d52d7c
• lora/adapter_model.safetensors · 0.06 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• lora/adapter_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• head.safetensors · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• hobson_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• provenance.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• tokenizer.json · 0.02 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• tokenizer_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd

• Base download includes the original multimodal checkpoint shard. This vision configuration retains the visual encoder, merger and decoder. Download bytes do not establish runtime working memory.
• FP32 CPU torso plus head, adapter, tokenizer, recurrent/attention state, packages and OS require more than just on-disk weight size. No generic token throughput claim is applicable.
• Vision loads the complete frozen visual encoder and merger plus language decoder, then CPU engine upcasts the torso to fp32. The docs0.66GB bf16 vision-tower increment is not a CPU required-memory measurement. Full peak/fit remains unknown.

Install the runtime

1. 1. Install the exact CPU source runtime

   Create a dedicated virtual environment and install the pinned source tree. Its pyproject declares torch >=2.7, transformers >=5.15,<6 for baseline text support, PEFT >=0.21 and related dependencies. No CUDA-specific extra is selected. Dependency versions are not a fully locked environment; retain the resolved freeze with any execution evidence. Vision explicitly requires transformers>=5.18,<6 and Pillow; the final freeze records these resolved versions.

   CommandCopy command

   sudo apt-get update
   sudo apt-get install -y python3 python3-venv python3-pip git curl ca-certificates
   mkdir -p "$HOME/strands-v19-vision-hosting"
   cd "$HOME/strands-v19-vision-hosting"
   python3 -m venv .venv
   source .venv/bin/activate
   python -m pip install --upgrade pip
   git clone https://github.com/strands-labs/strands-decider.git runtime
   git -C runtime checkout --detach 6039cf8b64c6074e75804000c07ac3d4cce795f7
   python -m pip install -e './runtime[cpu,vision]'
   python -m pip install "transformers>=5.18,<6" "pillow>=10"
   python -m pip freeze > installed-packages.txt

   Expected result: Source checkout is exactly pinned; installation succeeds and installed-packages.txt records resolved dependencies. No inference has run yet.

Download the model

1. 1. Freeze the complete v19 export and its documented base revision

   Download the full v19 snapshot into a local folder so the server never resolves checkpoint main. Prefetch the published base revision into the normal Hub cache. The base pin is explicitly inferred in provenance, not proof of the exact historical training revision. Keeping the full snapshot allows the publisher manifest verification.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   source .venv/bin/activate
   python - <<'PYDOWNLOAD'
   from huggingface_hub import snapshot_download
   snapshot_download(repo_id='StrandsAgents/strands-decider-2B-hobson-v19', revision='bb282d786bc251fd4e3068de3ada9ddbb38127cd', local_dir='./checkpoint')
   snapshot_download(repo_id='Qwen/Qwen3.5-2B-Base', revision='b1485b2fa6dfa1287294f269f5fb618e03d52d7c')
   PYDOWNLOAD
   python -m strands_decider.hf_export verify ./checkpoint

   Expected result: Complete checkpoint files match MANIFEST.sha256. Base cache resolves the explicit revision. Stop on missing adapter/head/config/tokenizer or verification failure.

2. 2. Verify critical adapter, head and base weight bytes

   The checkpoint manifest and explicit hashes bind the v19 adapter/head; verify the single base shard in the pinned Hub cache as well. No file is modified to substitute a different base.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   sha256sum --check <<'SHACHECK'
   701bdb895887097f7954ec7eb06f7937d3b790035b5462195eb26abd80a4aebc  checkpoint/lora/adapter_model.safetensors
   daad0727152b6185447cee36f78230c144312feb7b19f02749b19645238b5287  checkpoint/head.safetensors
   SHACHECK
   source .venv/bin/activate
   python - <<'PYBASE'
   from huggingface_hub import hf_hub_download
   import hashlib
   f=hf_hub_download(repo_id='Qwen/Qwen3.5-2B-Base',filename='model.safetensors-00001-of-00001.safetensors',revision='b1485b2fa6dfa1287294f269f5fb618e03d52d7c')
   h=hashlib.sha256()
   with open(f,'rb') as stream:
       for chunk in iter(lambda: stream.read(8*1024*1024), b''):
           h.update(chunk)
   assert h.hexdigest() == '928acbf11878c32185bbd863514d191769285065ab9ea14fbfe431303f5fdf2d', 'Base weight mismatch'
   print('Pinned base weight SHA256 OK')
   PYBASE

   Expected result: Adapter and head report OK; pinned base weight SHA256 OK. This verifies bytes, not runtime fit or accuracy.

3. 3. Fetch exact tiny visual smoke fixture

   Download the pinned official llama.cpp test image solely as a private test input; no photograph redistribution rights asserted. This is independent of the Strands model/runtime pins. Stop on mismatch.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   curl --fail --location --output fixture.jpg 'https://raw.githubusercontent.com/ggml-org/llama.cpp/620d2225cc436ab0b001e3200e590e3b286ef9ab/tools/mtmd/test-1.jpeg'
   sha256sum --check <<'SHA256'
   2dff664c0c8aaea18aff8cbe7e868845b775e90cdd7a0bac98df709b131deaa3  fixture.jpg
   SHA256

   Expected result: fixture.jpg: OK;124071 bytes. Image visibly shows The New York Times front page and MEN WALK ON MOON headline.

Configure

1. 1. Check base resolution and CPU settings

   The vision loader reads hobson_config.json for this legacy v19 export, resolves config.base_revision first or matching provenance.base_model_revision, and passes that revision to AutoConfig and Qwen3_5ForConditionalGeneration. It retains .model (visual encoder/merger and language_model), remaps the decoder-trained LoRA keys under language_model, restores the pointer head and freezes the vision tower. On CPU the engine upcasts the multimodal torso to FP32; readout is FP32. Confirm model/checkpoint/base identities before starting. Config num_slots=24 is metadata, not a pointer-head choice ceiling. OMP/MKL four-thread settings are allocations; use one client request at a time.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   source .venv/bin/activate
   python - <<'PYCHECK'
   import json
   from strands_decider.modeling import StrandsDeciderConfig,base_revision,config_path
   p='./checkpoint'
   c=StrandsDeciderConfig.from_json(config_path(p))
   assert c.base_model == 'Qwen/Qwen3.5-2B-Base'
   assert base_revision(p,c) == 'b1485b2fa6dfa1287294f269f5fb618e03d52d7c'
   assert c.max_length == 4096 and c.num_slots == 24 and c.use_lora
   print('v19 config and published inferred base revision match')
   PYCHECK

   Expected result: Config window 4096, num_slots metadata 24 and published inferred base revision match. This metadata check does not load model weights into inference.

Start the model

1. 1. Serve native image and typed text decisions on CPU

   Serve frozen v19 on CPU with native vision and strict4096-token window. CPU upcasts complete multimodal torso to fp32; readout already fp32. Native image path always uses prefix/suffix cache; --max-batch1 evaluates one question at a time and re-encodes image per chunk. Trusted loopback only; no authentication provided by this runtime.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   source .venv/bin/activate
   export OMP_NUM_THREADS=4 MKL_NUM_THREADS=4
   strands-decider serve ./checkpoint --host 127.0.0.1 --port 8000 --device cpu --model-name strands-v19 --strict-window --max-batch 1  --vision

   Expected result: Weights, adapter and head load; health becomes available on loopback. Stop if files, base identity or loading fail.

Verify the setup

1. 1. Verify health and all three decision primitives

   Use a second terminal on the server. No API key is required because the native server has no authentication. Do not expose it publicly. Health must identify this v19 setup, CPU, max_length 4096, vision true and prefix_cache true. The example checks an obvious billing triage case; probabilities/confidence are not guaranteed to equal the old model-card example.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   curl --fail http://127.0.0.1:8000/health
   cat > decision-request.json <<'REQUEST'
   {
     "state": "Help! My payouts have been failing for 3 days!",
     "questions": {
       "urgent": {
         "type": "noul",
         "instructions": "Does this convey urgency?"
       },
       "team": {
         "type": "choice",
         "instructions": "Which team should handle this?",
         "criteria": {
           "billing": "payout and payment problems",
           "sales": "buying a product",
           "retail": "physical store requests"
         }
       },
       "frustration": {
         "type": "score",
         "instructions": "How frustrated is the writer?",
         "criteria": [
           "calm",
           "frustrated",
           "very frustrated"
         ]
       }
     }
   }
   REQUEST
   curl --fail http://127.0.0.1:8000/v1/systemone -H 'Content-Type: application/json' --data-binary @decision-request.json --output decision-response.json
   cat decision-response.json

   Expected result: HTTP 200 response with model strands-v19 and answers.urgent.noul in [0,1], answers.team.choice naming a declared team with probabilities/confidence, and answers.frustration.score in [0,2] with level legend/probabilities/confidence. The obvious team should be billing; inspect it. No choices[0].message.content or generated-token claim is expected. Wrong labels require task validation, not blind automation.

2. 2. Check grounded image typed decisions

   Run on server host or its SSH shell. The raw JPEG is base64 encoded in images, not OpenAI content parts. Health must report vision:true. Expected grounded choice is moon_landing and newspaper noul P(true)>0.5, based on visual inspection of the fixture. These are author smoke-check criteria, not predicted numeric outputs. Legibility score is subjective: check its typed range/distribution, do not prescribe a score/confidence. No output prose is expected.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   source .venv/bin/activate
   python - <<'PYREQUEST'
   import pathlib,json,base64
   request={'state': 'Inspect the supplied newspaper front page. Classify only what the image visibly supports.', 'images': ['REPLACE_WITH_BASE64_FIXTURE'], 'questions': {'topic': {'type': 'choice', 'instructions': 'What is the main front-page topic?', 'criteria': {'moon_landing': 'Astronauts walking on the Moon', 'mars_landing': 'Astronauts walking on Mars', 'sports': 'A sports championship'}}, 'newspaper': {'type': 'noul', 'instructions': 'Is this a newspaper front page?'}, 'headline_legibility': {'type': 'score', 'instructions': 'How readable is the largest headline?', 'criteria': ['not readable', 'partly readable', 'clearly readable']}}}
   request['images']=[base64.b64encode(pathlib.Path('fixture.jpg').read_bytes()).decode()]
   pathlib.Path('image-request.json').write_text(json.dumps(request))
   PYREQUEST
   curl --fail http://127.0.0.1:8000/health
   curl --fail http://127.0.0.1:8000/v1/systemone -H 'Content-Type: application/json' --data-binary @image-request.json --output image-response.json
   python - <<'PYCHECK'
   import json,math
   r=json.load(open('image-response.json'));a=r['answers']
   assert a['topic']['type']=='choice' and a['topic']['choice']=='moon_landing'
   assert 0<=a['newspaper']['noul']<=1 and a['newspaper']['noul']>0.5
   assert a['headline_legibility']['type']=='score' and 0<=a['headline_legibility']['score']<=2
   for key in ['topic','headline_legibility']:
    p=a[key]['probabilities'];assert all(math.isfinite(v) and 0<=v<=1 for v in p.values())
    assert abs(sum(p.values())-1)<5e-4
   print('Grounded image smoke check passed; inspect full distributions in image-response.json')
   PYCHECK

   Expected result: Health HTTP200 with vision:true; successful native typed answers and passing assertions. Image rejection422, wrong topic, low newspaper probability or invalid schema fails the smoke check. Exact probability/confidence not promised.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• This is a typed decision model, not a general text-chat or token-generation endpoint. noul gives P(true); choice gives a named option; score gives an expectation over ordered level indices. Generated output tokens are not applicable.
• The frozen v19 export is a LoRA adapter plus pointer head over Qwen3.5-2B-Base. All three are required. The source loader preserves v19’s legacy config and reads the published inferred base pin; training hosts did not pin that base, so this is a reproducible published inference choice, not certain historical training provenance.
• Checkpoint max_length is 4096. Config num_slots=24 is retained metadata; this pointer-head checkpoint does not impose that fixed-slot choice ceiling. Public choice schema permits up to255 labels, while tokenized options/questions must fit the strict context window. Score rubrics use2–10 ascending levels. No different vision-specific option ceiling appears in the pinned vision path; strict-window overflow returns HTTP422.
• CPU execution upcasts its half-precision torso to FP32 and uses CPU-specific kernels; the head is FP32. No full process memory, per-request latency, throughput or cost was measured. Do not infer peak RAM from download bytes or transfer another version/device benchmark.
• Per-kind calibration temperatures are from v19 config (noul 0.9107136998460428, choice 0.734189596436441, score 1.32780942142348), not generation sampling settings. Calibration was fitted to held-out short tasks; measure your own traffic before trusting a confidence threshold.
• Publisher warns question wording is read less than state/options, long multi-step documents are weak, and score/noul transfer poorly to unfamiliar rubrics. Do not equate a high returned confidence with safety or correctness. Validate the intended task and failure handling.
• The current pinned source is newer than the original v19 card. No newer v1 benchmark/calibration results or exact old example probabilities are transferred. Dependency ranges are declared but not fully locked; preserve installed-packages.txt with actual run evidence.
• At most4 PNG/JPEG/WebP/GIF images. Source checks total pixel area width*height<=16777216 before decode; docs4096x4096 is shorthand and does not imply individual dimension limits. EXIF rotation, white transparency composition, RGB and bicubic downscale-only longest side448 precede pinned PIL image processing. A448px square is documented196 tokens; actual fixture count remains unknown.
• Images/state/questions share4096 window; strict-window rejects overflow422 rather than truncating. The image path always uses prefix cache regardless text-only no-prefix control. Full memory fit, CPU speed and vision accuracy are unknown; native probabilities cannot guarantee an image was read correctly.

Missing adapter/head, bad checksum or unexpected base model

Stop. Restore the complete exact v19 snapshot and manifest, including lora/, head, config, tokenizer and provenance. Verify critical SHA256 values and the inferred base revision. Do not serve just the adapter or silently fall back to main.

Import error or unknown Qwen3.5 class

Confirm pinned source checkout, active virtual environment and declared dependency constraints. Avoid assuming PyPI 0.1.0 includes current device extras. Keep the resolver log and pip freeze; do not substitute a generative chat runtime.

HTTP 422 or slow/out-of-memory request

Check the public schema limit of 255 choice labels, score 2–10 ordered levels, and tokenized image/state/questions within the strict 4096-token window. This pointer-head checkpoint does not impose its config num_slots=24 as a fixed-slot choice cap. Process one request at a time; free/provision resources or shorten input and record changes. No measured fit or speed is promised.

High-confidence wrong choice or unsuitable score

Check that state/options express the intended decision, simplify long multi-step documents and validate on representative labeled traffic. Adjust task design or train/calibrate separately; do not enable automatic action based solely on confidence.

Connection refused

Check foreground server logs, CPU health and loopback port 8000. Run the client on the same trusted desktop.

Stop the model

1. 1. Stop the server and release the model

   Press Ctrl+C in the foreground server terminal. Confirm localhost:8000 no longer serves decisions.

   Expected result: Server stops; loopback endpoint no longer responds.

Uninstall

1. 1. Remove only dedicated walkthrough files

   Back up wanted request/results and installed-packages.txt. Remove the dedicated folder after confirming its path. The shared Hugging Face base cache remains; list/manage it separately with the installed Hub CLI and remove only unwanted exact repositories/revisions, as other workloads may share it.

   CommandCopy command

   rm -r -- "$HOME/strands-v19-vision-hosting"

   Expected result: Dedicated runtime/venv/checkpoint files removed. Shared Hub cache and system packages are not blindly deleted.

Related guides

• https://opentools.ai/llms/strands-decider

Exact model and requirements

Artifact:

Strands Agents; Qwen base model / Strands Decider 2B hobson v19 base + LoRA + pointer head · checkpoint StrandsAgents/strands-decider-2B-hobson-v19

Publisher revision:

bb282d786bc251fd4e3068de3ada9ddbb38127cd

Artifact digest:

Composite checkpoint and separately pinned base have multiple hashes; exact component SHA256 values are pinned in auxiliary artifacts and recipe.

Download size:

4.3 GiB

Memory requirement (unknown):

Unknown — No exact full-process CPU peak for pinned base/adapter/head workload.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no model sharding. CPU-only torch execution; no accelerator or disk offload configured.

KV cache:

Torch runtime manages attention cache and recurrent state. Native vision path uses a prefix/suffix cache; exact state/cache precision and peak bytes are not separately measured.

Workload:

4,096 token context; Unknown; Not applicable; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Exact v19 exported tokenizer · bb282d786bc251fd4e3068de3ada9ddbb38127cd

Sampling / reasoning:

No generative sampler. Frozen per-kind calibration temperatures from v19 config; typed noul/choice/score readouts. No generated thinking or prose. Native vision prefix/suffix path uses a cache unconditionally; --max-batch1 re-encodes the image for each question chunk. Request contains image plus typed questions.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Strands v19 and its runtime declare Apache License 2.0; the pinned Qwen base is also Apache 2.0. Read and retain applicable license, copyright, attribution and modification notices when distributing covered work. Dataset attribution in the v19 license does not make unrelated dataset uses automatically permitted.

• Qwen base weight shard · 4.24 GiB · revision b1485b2fa6dfa1287294f269f5fb618e03d52d7c
• Qwen base config · 0 GiB · revision b1485b2fa6dfa1287294f269f5fb618e03d52d7c
• lora/adapter_model.safetensors · 0.06 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• lora/adapter_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• head.safetensors · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• hobson_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• provenance.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• tokenizer.json · 0.02 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd
• tokenizer_config.json · 0 GiB · revision bb282d786bc251fd4e3068de3ada9ddbb38127cd

• Base download includes the original multimodal checkpoint shard. This vision configuration retains the visual encoder, merger and decoder. Download bytes do not establish runtime working memory.
• FP32 CPU torso plus head, adapter, tokenizer, recurrent/attention state, packages and OS require more than just on-disk weight size. No generic token throughput claim is applicable.
• Vision loads the complete frozen visual encoder and merger plus language decoder, then CPU engine upcasts the torso to fp32. The docs0.66GB bf16 vision-tower increment is not a CPU required-memory measurement. Full peak/fit remains unknown.

Install the runtime

1. 1. Install the exact CPU source runtime

   Create a dedicated virtual environment and install the pinned source tree. Its pyproject declares torch >=2.7, transformers >=5.15,<6 for baseline text support, PEFT >=0.21 and related dependencies. No CUDA-specific extra is selected. Dependency versions are not a fully locked environment; retain the resolved freeze with any execution evidence. Vision explicitly requires transformers>=5.18,<6 and Pillow; the final freeze records these resolved versions.

   CommandCopy command

   sudo apt-get update
   sudo apt-get install -y python3 python3-venv python3-pip git curl ca-certificates
   mkdir -p "$HOME/strands-v19-vision-hosting"
   cd "$HOME/strands-v19-vision-hosting"
   python3 -m venv .venv
   source .venv/bin/activate
   python -m pip install --upgrade pip
   git clone https://github.com/strands-labs/strands-decider.git runtime
   git -C runtime checkout --detach 6039cf8b64c6074e75804000c07ac3d4cce795f7
   python -m pip install -e './runtime[cpu,vision]'
   python -m pip install "transformers>=5.18,<6" "pillow>=10"
   python -m pip freeze > installed-packages.txt

   Expected result: Source checkout is exactly pinned; installation succeeds and installed-packages.txt records resolved dependencies. No inference has run yet.

Download the model

1. 1. Freeze the complete v19 export and its documented base revision

   Download the full v19 snapshot into a local folder so the server never resolves checkpoint main. Prefetch the published base revision into the normal Hub cache. The base pin is explicitly inferred in provenance, not proof of the exact historical training revision. Keeping the full snapshot allows the publisher manifest verification.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   source .venv/bin/activate
   python - <<'PYDOWNLOAD'
   from huggingface_hub import snapshot_download
   snapshot_download(repo_id='StrandsAgents/strands-decider-2B-hobson-v19', revision='bb282d786bc251fd4e3068de3ada9ddbb38127cd', local_dir='./checkpoint')
   snapshot_download(repo_id='Qwen/Qwen3.5-2B-Base', revision='b1485b2fa6dfa1287294f269f5fb618e03d52d7c')
   PYDOWNLOAD
   python -m strands_decider.hf_export verify ./checkpoint

   Expected result: Complete checkpoint files match MANIFEST.sha256. Base cache resolves the explicit revision. Stop on missing adapter/head/config/tokenizer or verification failure.

2. 2. Verify critical adapter, head and base weight bytes

   The checkpoint manifest and explicit hashes bind the v19 adapter/head; verify the single base shard in the pinned Hub cache as well. No file is modified to substitute a different base.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   sha256sum --check <<'SHACHECK'
   701bdb895887097f7954ec7eb06f7937d3b790035b5462195eb26abd80a4aebc  checkpoint/lora/adapter_model.safetensors
   daad0727152b6185447cee36f78230c144312feb7b19f02749b19645238b5287  checkpoint/head.safetensors
   SHACHECK
   source .venv/bin/activate
   python - <<'PYBASE'
   from huggingface_hub import hf_hub_download
   import hashlib
   f=hf_hub_download(repo_id='Qwen/Qwen3.5-2B-Base',filename='model.safetensors-00001-of-00001.safetensors',revision='b1485b2fa6dfa1287294f269f5fb618e03d52d7c')
   h=hashlib.sha256()
   with open(f,'rb') as stream:
       for chunk in iter(lambda: stream.read(8*1024*1024), b''):
           h.update(chunk)
   assert h.hexdigest() == '928acbf11878c32185bbd863514d191769285065ab9ea14fbfe431303f5fdf2d', 'Base weight mismatch'
   print('Pinned base weight SHA256 OK')
   PYBASE

   Expected result: Adapter and head report OK; pinned base weight SHA256 OK. This verifies bytes, not runtime fit or accuracy.

3. 3. Fetch exact tiny visual smoke fixture

   Download the pinned official llama.cpp test image solely as a private test input; no photograph redistribution rights asserted. This is independent of the Strands model/runtime pins. Stop on mismatch.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   curl --fail --location --output fixture.jpg 'https://raw.githubusercontent.com/ggml-org/llama.cpp/620d2225cc436ab0b001e3200e590e3b286ef9ab/tools/mtmd/test-1.jpeg'
   sha256sum --check <<'SHA256'
   2dff664c0c8aaea18aff8cbe7e868845b775e90cdd7a0bac98df709b131deaa3  fixture.jpg
   SHA256

   Expected result: fixture.jpg: OK;124071 bytes. Image visibly shows The New York Times front page and MEN WALK ON MOON headline.

Configure

1. 1. Check base resolution and CPU settings

   The vision loader reads hobson_config.json for this legacy v19 export, resolves config.base_revision first or matching provenance.base_model_revision, and passes that revision to AutoConfig and Qwen3_5ForConditionalGeneration. It retains .model (visual encoder/merger and language_model), remaps the decoder-trained LoRA keys under language_model, restores the pointer head and freezes the vision tower. On CPU the engine upcasts the multimodal torso to FP32; readout is FP32. Confirm model/checkpoint/base identities before starting. Config num_slots=24 is metadata, not a pointer-head choice ceiling. OMP/MKL four-thread settings are allocations; use one client request at a time.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   source .venv/bin/activate
   python - <<'PYCHECK'
   import json
   from strands_decider.modeling import StrandsDeciderConfig,base_revision,config_path
   p='./checkpoint'
   c=StrandsDeciderConfig.from_json(config_path(p))
   assert c.base_model == 'Qwen/Qwen3.5-2B-Base'
   assert base_revision(p,c) == 'b1485b2fa6dfa1287294f269f5fb618e03d52d7c'
   assert c.max_length == 4096 and c.num_slots == 24 and c.use_lora
   print('v19 config and published inferred base revision match')
   PYCHECK

   Expected result: Config window 4096, num_slots metadata 24 and published inferred base revision match. This metadata check does not load model weights into inference.

Start the model

1. 1. Serve native image and typed text decisions on CPU

   Serve frozen v19 on CPU with native vision and strict4096-token window. CPU upcasts complete multimodal torso to fp32; readout already fp32. Native image path always uses prefix/suffix cache; --max-batch1 evaluates one question at a time and re-encodes image per chunk. Trusted loopback only; no authentication provided by this runtime.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   source .venv/bin/activate
   export OMP_NUM_THREADS=4 MKL_NUM_THREADS=4
   strands-decider serve ./checkpoint --host 127.0.0.1 --port 8000 --device cpu --model-name strands-v19 --strict-window --max-batch 1  --vision

   Expected result: Weights, adapter and head load; health becomes available on loopback. Stop if files, base identity or loading fail.

Verify the setup

1. 1. Verify health and all three decision primitives

   Use a second terminal on the server. No API key is required because the native server has no authentication. Do not expose it publicly. Health must identify this v19 setup, CPU, max_length 4096, vision true and prefix_cache true. The example checks an obvious billing triage case; probabilities/confidence are not guaranteed to equal the old model-card example.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   curl --fail http://127.0.0.1:8000/health
   cat > decision-request.json <<'REQUEST'
   {
     "state": "Help! My payouts have been failing for 3 days!",
     "questions": {
       "urgent": {
         "type": "noul",
         "instructions": "Does this convey urgency?"
       },
       "team": {
         "type": "choice",
         "instructions": "Which team should handle this?",
         "criteria": {
           "billing": "payout and payment problems",
           "sales": "buying a product",
           "retail": "physical store requests"
         }
       },
       "frustration": {
         "type": "score",
         "instructions": "How frustrated is the writer?",
         "criteria": [
           "calm",
           "frustrated",
           "very frustrated"
         ]
       }
     }
   }
   REQUEST
   curl --fail http://127.0.0.1:8000/v1/systemone -H 'Content-Type: application/json' --data-binary @decision-request.json --output decision-response.json
   cat decision-response.json

   Expected result: HTTP 200 response with model strands-v19 and answers.urgent.noul in [0,1], answers.team.choice naming a declared team with probabilities/confidence, and answers.frustration.score in [0,2] with level legend/probabilities/confidence. The obvious team should be billing; inspect it. No choices[0].message.content or generated-token claim is expected. Wrong labels require task validation, not blind automation.

2. 2. Check grounded image typed decisions

   Run on server host or its SSH shell. The raw JPEG is base64 encoded in images, not OpenAI content parts. Health must report vision:true. Expected grounded choice is moon_landing and newspaper noul P(true)>0.5, based on visual inspection of the fixture. These are author smoke-check criteria, not predicted numeric outputs. Legibility score is subjective: check its typed range/distribution, do not prescribe a score/confidence. No output prose is expected.

   CommandCopy command

   cd "$HOME/strands-v19-vision-hosting"
   source .venv/bin/activate
   python - <<'PYREQUEST'
   import pathlib,json,base64
   request={'state': 'Inspect the supplied newspaper front page. Classify only what the image visibly supports.', 'images': ['REPLACE_WITH_BASE64_FIXTURE'], 'questions': {'topic': {'type': 'choice', 'instructions': 'What is the main front-page topic?', 'criteria': {'moon_landing': 'Astronauts walking on the Moon', 'mars_landing': 'Astronauts walking on Mars', 'sports': 'A sports championship'}}, 'newspaper': {'type': 'noul', 'instructions': 'Is this a newspaper front page?'}, 'headline_legibility': {'type': 'score', 'instructions': 'How readable is the largest headline?', 'criteria': ['not readable', 'partly readable', 'clearly readable']}}}
   request['images']=[base64.b64encode(pathlib.Path('fixture.jpg').read_bytes()).decode()]
   pathlib.Path('image-request.json').write_text(json.dumps(request))
   PYREQUEST
   curl --fail http://127.0.0.1:8000/health
   curl --fail http://127.0.0.1:8000/v1/systemone -H 'Content-Type: application/json' --data-binary @image-request.json --output image-response.json
   python - <<'PYCHECK'
   import json,math
   r=json.load(open('image-response.json'));a=r['answers']
   assert a['topic']['type']=='choice' and a['topic']['choice']=='moon_landing'
   assert 0<=a['newspaper']['noul']<=1 and a['newspaper']['noul']>0.5
   assert a['headline_legibility']['type']=='score' and 0<=a['headline_legibility']['score']<=2
   for key in ['topic','headline_legibility']:
    p=a[key]['probabilities'];assert all(math.isfinite(v) and 0<=v<=1 for v in p.values())
    assert abs(sum(p.values())-1)<5e-4
   print('Grounded image smoke check passed; inspect full distributions in image-response.json')
   PYCHECK

   Expected result: Health HTTP200 with vision:true; successful native typed answers and passing assertions. Image rejection422, wrong topic, low newspaper probability or invalid schema fails the smoke check. Exact probability/confidence not promised.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• This is a typed decision model, not a general text-chat or token-generation endpoint. noul gives P(true); choice gives a named option; score gives an expectation over ordered level indices. Generated output tokens are not applicable.
• The frozen v19 export is a LoRA adapter plus pointer head over Qwen3.5-2B-Base. All three are required. The source loader preserves v19’s legacy config and reads the published inferred base pin; training hosts did not pin that base, so this is a reproducible published inference choice, not certain historical training provenance.
• Checkpoint max_length is 4096. Config num_slots=24 is retained metadata; this pointer-head checkpoint does not impose that fixed-slot choice ceiling. Public choice schema permits up to255 labels, while tokenized options/questions must fit the strict context window. Score rubrics use2–10 ascending levels. No different vision-specific option ceiling appears in the pinned vision path; strict-window overflow returns HTTP422.
• CPU execution upcasts its half-precision torso to FP32 and uses CPU-specific kernels; the head is FP32. No full process memory, per-request latency, throughput or cost was measured. Do not infer peak RAM from download bytes or transfer another version/device benchmark.
• Per-kind calibration temperatures are from v19 config (noul 0.9107136998460428, choice 0.734189596436441, score 1.32780942142348), not generation sampling settings. Calibration was fitted to held-out short tasks; measure your own traffic before trusting a confidence threshold.
• Publisher warns question wording is read less than state/options, long multi-step documents are weak, and score/noul transfer poorly to unfamiliar rubrics. Do not equate a high returned confidence with safety or correctness. Validate the intended task and failure handling.
• The current pinned source is newer than the original v19 card. No newer v1 benchmark/calibration results or exact old example probabilities are transferred. Dependency ranges are declared but not fully locked; preserve installed-packages.txt with actual run evidence.
• At most4 PNG/JPEG/WebP/GIF images. Source checks total pixel area width*height<=16777216 before decode; docs4096x4096 is shorthand and does not imply individual dimension limits. EXIF rotation, white transparency composition, RGB and bicubic downscale-only longest side448 precede pinned PIL image processing. A448px square is documented196 tokens; actual fixture count remains unknown.
• Images/state/questions share4096 window; strict-window rejects overflow422 rather than truncating. The image path always uses prefix cache regardless text-only no-prefix control. Full memory fit, CPU speed and vision accuracy are unknown; native probabilities cannot guarantee an image was read correctly.

Missing adapter/head, bad checksum or unexpected base model

Stop. Restore the complete exact v19 snapshot and manifest, including lora/, head, config, tokenizer and provenance. Verify critical SHA256 values and the inferred base revision. Do not serve just the adapter or silently fall back to main.

Import error or unknown Qwen3.5 class

Confirm pinned source checkout, active virtual environment and declared dependency constraints. Avoid assuming PyPI 0.1.0 includes current device extras. Keep the resolver log and pip freeze; do not substitute a generative chat runtime.

HTTP 422 or slow/out-of-memory request

Check the public schema limit of 255 choice labels, score 2–10 ordered levels, and tokenized image/state/questions within the strict 4096-token window. This pointer-head checkpoint does not impose its config num_slots=24 as a fixed-slot choice cap. Process one request at a time; free/provision resources or shorten input and record changes. No measured fit or speed is promised.

High-confidence wrong choice or unsuitable score

Check that state/options express the intended decision, simplify long multi-step documents and validate on representative labeled traffic. Adjust task design or train/calibrate separately; do not enable automatic action based solely on confidence.

Connection refused

Check foreground server logs, CPU health and loopback port 8000. Verify the SSH tunnel to the intended VM; do not expose the unauthenticated port.

Stop the model

1. 1. Stop the server and release the model

   Press Ctrl+C in the foreground server terminal. Close the SSH tunnel with Ctrl+C; stopping only this process does not stop EC2 billing.

   Expected result: Server stops; loopback endpoint no longer responds.

Uninstall

1. 1. Remove only dedicated walkthrough files

   Back up wanted request/results and installed-packages.txt. Remove the dedicated folder after confirming its path. The shared Hugging Face base cache remains; list/manage it separately with the installed Hub CLI and remove only unwanted exact repositories/revisions, as other workloads may share it.

   CommandCopy command

   rm -r -- "$HOME/strands-v19-vision-hosting"

   Expected result: Dedicated runtime/venv/checkpoint files removed. Shared Hub cache and system packages are not blindly deleted.

Shut down cloud resources

1. 1. End cloud billing resources

   In EC2 us-east-1, stop for later reuse or terminate this exact VM. Review Delete on termination for its root volume. Delete unwanted detached EBS volumes, snapshots and reserved/Elastic/public IP resources. Stopped VM storage and retained resources may still bill. Check Bills/Cost Explorer later for final usage.

   Expected result: Exact instance is stopped/terminated as intended; no unwanted attached/retained billed resources remain. Instance compute stops only after its lifecycle transition, not when llama-server stops.

Related guides

• https://opentools.ai/llms/strands-decider