Reka Flash 3

Exact model and requirements

Artifact:

Unsloth (GGUF derivative); Reka AI (native model) / Reka Flash 3 Q4_K_M, Unsloth derivative · checkpoint unsloth/reka-flash-3-GGUF

Publisher revision:

3d2f2c32bcbc0393ed6d47f37b50ebd2a5f41f3e

Artifact digest:

sha256:abc4348fe1e6bf5d5527aad5672a05732f29022a199c6bf6f16fba992855c367

Download size:

12.68 GiB

Memory requirement (unknown):

Unknown — No exact CPU process peak for this 21B dense GGUF, 8K context and native reasoning workload.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no tensor/pipeline sharding. CPU-only execution; --device none and zero GPU layers. No disk or accelerator offload.

KV cache:

F16 K and V attention cache

Workload:

8,192 token context; Unknown; Unknown; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Embedded cl100k_base-compatible GGUF tokenizer and separately extracted native Reka chat template · GGUF 3d2f2c32bcbc0393ed6d47f37b50ebd2a5f41f3e; native template 8d251ecb02eca9f0f4aa052b29ef922628d2806f; derived SHA256 1e330d18b71022b555c66089bebea6e4977d5fbc9a39534bb2e6e3d1af9829ed

Sampling / reasoning:

Native generation_config: temperature 0.6, top_k 1024 and top_p 0.95. Author disables extra min_p filtering (0), sets repeat_penalty 1 and frequency/presence penalties 0. Request max_tokens 4096 includes reasoning and final content. Reka natively thinks using <reasoning> and </reasoning>. Raw-content output preserves those tags; no separate reasoning parser or disabled-thinking switch. This guide does not implement publisher budget forcing: a total output cap does not force a closing tag or final answer. Stop on truncation or missing final content.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Reka AI native model-card metadata and the Unsloth derivative metadata both declare Apache-2.0; the derivative explicitly names RekaAI/reka-flash-3. The native repository has no standalone root LICENSE at this revision (404). Preserve license, attribution/copyright and required notices, identify modifications and review patent/NOTICE conditions for redistribution. This is an Unsloth GGUF, not a Reka-published binary. Public self-host download does not require a provider inference subscription.

• tokenizer_config.json containing the native chat template · 0 GiB · revision 8d251ecb02eca9f0f4aa052b29ef922628d2806f

• The full 13,610,362,464-byte dense-model artifact must be resident with runtime state. File size alone does not establish working RAM.
• Attention cache, reasoning output, compute buffers, runtime, OS and setup add requirements. 32 GiB RAM, 40 GiB disk and four CPU threads are author allocations, not measured minimums or fit/speed guarantees.

Install the runtime

1. 1. Build the pinned CPU server

   Install build dependencies and compile the exact official source revision. These packages are system shared dependencies; no CUDA/GPU driver is required. GGML_NATIVE=ON selects -march=native for this GNU/Linux x86 build, compiling for the host CPU. Build and run on the same machine; no AVX-512 requirement is imposed on the i5-12400.

   CommandCopy command

   sudo apt-get update
   sudo apt-get install -y build-essential cmake git curl ca-certificates openssl python3
   mkdir -p "$HOME/reka-flash3-hosting/models"
   cd "$HOME/reka-flash3-hosting"
   git clone https://github.com/ggml-org/llama.cpp.git
   cd llama.cpp
   git checkout --detach 620d2225cc436ab0b001e3200e590e3b286ef9ab
   cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_METAL=OFF -DGGML_NATIVE=ON
   cmake --build build --config Release -j 4
   ./build/bin/llama-server --version

   Expected result: Build succeeds and server version identifies pinned source. Stop if checkout or build fails.

Download the model

1. 1. Download the exact language derivative and original tokenizer configuration

   Verify immutable GGUF bytes and the pinned publisher JSON. The native template lives inside tokenizer_config.json; a missing standalone chat_template.jinja is expected. Stop on either checksum mismatch.

   CommandCopy command

   cd "$HOME/reka-flash3-hosting"
   curl --fail --location --output models/reka-flash-3-Q4_K_M.gguf 'https://huggingface.co/unsloth/reka-flash-3-GGUF/resolve/3d2f2c32bcbc0393ed6d47f37b50ebd2a5f41f3e/reka-flash-3-Q4_K_M.gguf'
   curl --fail --location --output models/tokenizer_config.json 'https://huggingface.co/RekaAI/reka-flash-3/raw/8d251ecb02eca9f0f4aa052b29ef922628d2806f/tokenizer_config.json'
   sha256sum --check <<'SHA256'
   abc4348fe1e6bf5d5527aad5672a05732f29022a199c6bf6f16fba992855c367  models/reka-flash-3-Q4_K_M.gguf
   e4bc8902a08408876a5cb3a9494529b99f5c5ba895111ac9faa3d5507dd101d9  models/tokenizer_config.json
   SHA256

   Expected result: Both files report OK. GGUF is 13,610,362,464 bytes; native tokenizer configuration is 5,777 bytes. No projector or other model size is downloaded.

Configure

1. 1. Create a private service credential

   Generate this only on your own host when following the guide. It is a local service password, not an OpenAI/provider credential. Keep it private and never paste it into tickets. Create it only when absent; preserve an existing key so a repeated setup does not silently rotate access. It is not created during this documentation run.

   CommandCopy command

   cd "$HOME/reka-flash3-hosting"
   umask 077
   if [ ! -e .api-key ]; then
     openssl rand -hex 32 > .api-key
   fi
   test -s .api-key || { echo 'Existing .api-key is empty; restore it before continuing.'; exit 1; }
   chmod 600 .api-key

   Expected result: A non-empty private .api-key file exists; existing key preserved. No secret is printed.

2. 2. Extract and verify the native chat template verbatim

   Decode the verified UTF-8 JSON, select its chat_template string and write those exact UTF-8 bytes without stripping whitespace or adding a newline. This is a documented derivation, not an independently hosted template. Compare the file back to the embedded string and its expected digest.

   CommandCopy command

   cd "$HOME/reka-flash3-hosting"
   python3 - <<'PYTEMPLATE'
   import json,pathlib,hashlib
   config=json.loads(pathlib.Path('models/tokenizer_config.json').read_text(encoding='utf-8'))
   text=config['chat_template']
   assert isinstance(text,str) and text
   raw=text.encode('utf-8')
   assert len(raw)==835 and hashlib.sha256(raw).hexdigest()=='1e330d18b71022b555c66089bebea6e4977d5fbc9a39534bb2e6e3d1af9829ed'
   pathlib.Path('models/reka-chat.jinja').write_bytes(raw)
   assert pathlib.Path('models/reka-chat.jinja').read_bytes()==config['chat_template'].encode('utf-8')
   print('Native template extracted verbatim and hash verified')
   PYTEMPLATE

   Expected result: Native template extracted verbatim and hash verified. The 835-byte template formats human:/assistant: rounds, joins a leading system message into the first user round, and strips completed prior-round reasoning traces.

Start the model

1. 1. Start one bounded native reasoning sequence

   Use the exact native template and raw content to retain <reasoning> tags. 8,192 total context includes prompt and output; reserve up to 4,096 generated tokens for both reasoning and final answer. This is a reduced smoke workload below the native 32,768 positions, not a full-capacity or successful-completion claim. No budget forcing or early </reasoning> injection is implemented.

   CommandCopy command

   cd "$HOME/reka-flash3-hosting"
   export REKA_FLASH3_SERVICE_KEY="$(cat .api-key)"
   ./llama.cpp/build/bin/llama-server -m models/reka-flash-3-Q4_K_M.gguf --alias reka-flash3 --host 127.0.0.1 --port 8080 --api-key "$REKA_FLASH3_SERVICE_KEY" --device none -ngl 0 -t 4 -c 8192 -np 1 --cache-type-k f16 --cache-type-v f16 --jinja --chat-template-file models/reka-chat.jinja --skip-chat-parsing --reasoning-format none --temp 0.6 --top-k 1024 --top-p 0.95 --min-p 0 --repeat-penalty 1 --frequency-penalty 0 --presence-penalty 0

   Expected result: Model initializes and loopback listener starts. Stop on tokenizer/architecture/template errors. The service key in process arguments may be visible to local/privileged users; use a trusted host.

Verify the setup

1. 1. Verify readiness, authenticated reasoning output and stop conditions

   Run in a second terminal on the server or its SSH shell. Explicit stop strings implement publisher guidance. Inspect raw choices[0].message.content for a completed </reasoning> block followed by a final answer of 36 counters; exact prose and token counts are not promised. A length finish, unclosed reasoning or absent/wrong final answer fails the smoke check. The cap does not guarantee reasoning completion.

   CommandCopy command

   cd "$HOME/reka-flash3-hosting"
   export REKA_FLASH3_SERVICE_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8080/health
   curl --fail http://127.0.0.1:8080/v1/chat/completions -H "Authorization: Bearer $REKA_FLASH3_SERVICE_KEY" -H 'Content-Type: application/json' --data-binary '{"model":"reka-flash3","messages":[{"role":"user","content":"A box contains 12 red counters, 18 blue counters and 6 green counters. How many counters are in the box? Think carefully and give a short final answer using only these numbers."}],"max_tokens":4096,"temperature":0.6,"top_k":1024,"top_p":0.95,"min_p":0,"repeat_penalty":1,"frequency_penalty":0,"presence_penalty":0,"stop":["<sep>","<|endoftext|>"]}' --output reasoning-response.json
   cat reasoning-response.json
   curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8080/v1/chat/completions -H 'Content-Type: application/json' --data-binary '{"model":"reka-flash3","messages":[{"role":"user","content":"Hello"}]}'

   Expected result: Ready health returns HTTP 200; authenticated generation returns nonempty raw content with completed native reasoning and a correct final total of 36, finish_reason not length. Missing service key returns HTTP 401. These are reader pass criteria, not observed inference results.

Performance evidence

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

Limits and troubleshooting

• Documented configuration only. No weights, installation/build, inference, fit, speed or quality observation was executed by the author.
• This is a dense Llama-compatible text model: 44 layers, hidden size 6,144, 64 attention heads, eight KV heads and head dimension 96. Native maximum positions 32,768 do not establish full-context fit in this 8K recipe.
• The publisher says Reka Flash thinks using <reasoning> tags and describes optional budget forcing by injecting a closing tag. This guide only caps total generated output at 4,096; it does not implement that technique or guarantee a final answer. Reasoning and final answer share the cap.
• Raw-content output preserves the native tags; do not expect a separate reasoning_content field. History formatting uses the pinned native template, which removes completed prior assistant reasoning traces.
• Stop on the literal <sep> string or <|endoftext|>; both are passed explicitly in the validation request. Do not replace native human:/assistant: formatting with a generic Llama instruction template.
• Native generation_config supplies temperature 0.6, top_k 1024 and top_p 0.95. Extra min-p/repetition/frequency/presence controls in this guide are author choices, not an additional publisher preset.
• The publisher recommends English use, warns about knowledge-intensive tasks and limited alignment/persona training. The supplied arithmetic task checks a narrow functional path, not general reliability, safety, coding or multilingual performance.
• The native/derivative metadata declare Apache-2.0. The Unsloth card names the Reka base without establishing an exact upstream conversion revision. Runnable derivative bytes and the independent publisher template are pinned separately.
• System packages resolve at installation. Preserve tool versions and build logs; GGML_NATIVE means build and run on the same CPU.
• 32 GiB RAM, 40 GiB disk and four threads are allocations, not measured minima or confirmed fit. Loopback/private bearer access assumes a trusted host; shared/public serving needs separate TLS, isolation and operations.
• Two guides do not establish complete model coverage. Other useful hardware/OS/runtime/full-context/budget-forcing/coding workflows remain unqualified, not excluded.

Checksum, tokenizer or template error

Stop. Confirm exact Unsloth file digest and native tokenizer JSON revision/hash. Re-extract chat_template verbatim and verify its derived SHA. Use the pinned native Llama runtime, not a generic prompt substitution.

Length finish, unclosed reasoning or no final answer

Treat the smoke check as failed. The reduced output budget is not guaranteed to complete reasoning. Record any changed output/context workload; do not silently cut tags or call a reasoning-only trace a final answer.

Memory pressure or context overflow

Stop and inspect host resources. Free/provision memory or explicitly reduce workload; file size does not prove fit. Preserve reasoning/output budget choices in any new configuration.

HTTP 401 or connection refused

Check foreground startup, loopback port 8080 and matching private service key. For cloud, verify the SSH tunnel points at the intended VM; keep public 8080 ingress closed.

Stop the model

1. 1. Stop the service

   Press Ctrl+C in server terminal. Unset key in each shell. Confirm this desktop localhost:8080 no longer serves the model.

   CommandCopy command

   unset REKA_FLASH3_SERVICE_KEY

   Expected result: Server stops and localhost:8080 stops serving.

Uninstall

1. 1. Remove dedicated model, build and credential files

   Back up wanted logs/files first. Remove only this walkthrough directory after confirming its path. Shared apt packages can remain; do not remove other workloads.

   CommandCopy command

   rm -r -- "$HOME/reka-flash3-hosting"

   Expected result: Dedicated files and local service credential removed; shared system dependencies remain.

Related guides

• https://opentools.ai/llms/reka-flash-3

Exact model and requirements

Artifact:

Unsloth (GGUF derivative); Reka AI (native model) / Reka Flash 3 Q4_K_M, Unsloth derivative · checkpoint unsloth/reka-flash-3-GGUF

Publisher revision:

3d2f2c32bcbc0393ed6d47f37b50ebd2a5f41f3e

Artifact digest:

sha256:abc4348fe1e6bf5d5527aad5672a05732f29022a199c6bf6f16fba992855c367

Download size:

12.68 GiB

Memory requirement (unknown):

Unknown — No exact CPU process peak for this 21B dense GGUF, 8K context and native reasoning workload.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no tensor/pipeline sharding. CPU-only execution; --device none and zero GPU layers. No disk or accelerator offload.

KV cache:

F16 K and V attention cache

Workload:

8,192 token context; Unknown; Unknown; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Embedded cl100k_base-compatible GGUF tokenizer and separately extracted native Reka chat template · GGUF 3d2f2c32bcbc0393ed6d47f37b50ebd2a5f41f3e; native template 8d251ecb02eca9f0f4aa052b29ef922628d2806f; derived SHA256 1e330d18b71022b555c66089bebea6e4977d5fbc9a39534bb2e6e3d1af9829ed

Sampling / reasoning:

Native generation_config: temperature 0.6, top_k 1024 and top_p 0.95. Author disables extra min_p filtering (0), sets repeat_penalty 1 and frequency/presence penalties 0. Request max_tokens 4096 includes reasoning and final content. Reka natively thinks using <reasoning> and </reasoning>. Raw-content output preserves those tags; no separate reasoning parser or disabled-thinking switch. This guide does not implement publisher budget forcing: a total output cap does not force a closing tag or final answer. Stop on truncation or missing final content.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Reka AI native model-card metadata and the Unsloth derivative metadata both declare Apache-2.0; the derivative explicitly names RekaAI/reka-flash-3. The native repository has no standalone root LICENSE at this revision (404). Preserve license, attribution/copyright and required notices, identify modifications and review patent/NOTICE conditions for redistribution. This is an Unsloth GGUF, not a Reka-published binary. Public self-host download does not require a provider inference subscription.

• tokenizer_config.json containing the native chat template · 0 GiB · revision 8d251ecb02eca9f0f4aa052b29ef922628d2806f

• The full 13,610,362,464-byte dense-model artifact must be resident with runtime state. File size alone does not establish working RAM.
• Attention cache, reasoning output, compute buffers, runtime, OS and setup add requirements. 32 GiB RAM, 40 GiB disk and four CPU threads are author allocations, not measured minimums or fit/speed guarantees.

Install the runtime

1. 1. Build the pinned CPU server

   Install build dependencies and compile the exact official source revision. These packages are system shared dependencies; no CUDA/GPU driver is required. GGML_NATIVE=ON selects -march=native for this GNU/Linux x86 build, compiling for the host CPU. Build and run on the same machine; the binary uses the instruction set exposed by this r7i.xlarge guest. Do not copy it to a different instance family without rebuilding.

   CommandCopy command

   sudo apt-get update
   sudo apt-get install -y build-essential cmake git curl ca-certificates openssl python3
   mkdir -p "$HOME/reka-flash3-hosting/models"
   cd "$HOME/reka-flash3-hosting"
   git clone https://github.com/ggml-org/llama.cpp.git
   cd llama.cpp
   git checkout --detach 620d2225cc436ab0b001e3200e590e3b286ef9ab
   cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_METAL=OFF -DGGML_NATIVE=ON
   cmake --build build --config Release -j 4
   ./build/bin/llama-server --version

   Expected result: Build succeeds and server version identifies pinned source. Stop if checkout or build fails.

Download the model

1. 1. Download the exact language derivative and original tokenizer configuration

   Verify immutable GGUF bytes and the pinned publisher JSON. The native template lives inside tokenizer_config.json; a missing standalone chat_template.jinja is expected. Stop on either checksum mismatch.

   CommandCopy command

   cd "$HOME/reka-flash3-hosting"
   curl --fail --location --output models/reka-flash-3-Q4_K_M.gguf 'https://huggingface.co/unsloth/reka-flash-3-GGUF/resolve/3d2f2c32bcbc0393ed6d47f37b50ebd2a5f41f3e/reka-flash-3-Q4_K_M.gguf'
   curl --fail --location --output models/tokenizer_config.json 'https://huggingface.co/RekaAI/reka-flash-3/raw/8d251ecb02eca9f0f4aa052b29ef922628d2806f/tokenizer_config.json'
   sha256sum --check <<'SHA256'
   abc4348fe1e6bf5d5527aad5672a05732f29022a199c6bf6f16fba992855c367  models/reka-flash-3-Q4_K_M.gguf
   e4bc8902a08408876a5cb3a9494529b99f5c5ba895111ac9faa3d5507dd101d9  models/tokenizer_config.json
   SHA256

   Expected result: Both files report OK. GGUF is 13,610,362,464 bytes; native tokenizer configuration is 5,777 bytes. No projector or other model size is downloaded.

Configure

1. 1. Create a private service credential

   Generate this only on your own host when following the guide. It is a local service password, not an OpenAI/provider credential. Keep it private and never paste it into tickets. Create it only when absent; preserve an existing key so a repeated setup does not silently rotate access. It is not created during this documentation run.

   CommandCopy command

   cd "$HOME/reka-flash3-hosting"
   umask 077
   if [ ! -e .api-key ]; then
     openssl rand -hex 32 > .api-key
   fi
   test -s .api-key || { echo 'Existing .api-key is empty; restore it before continuing.'; exit 1; }
   chmod 600 .api-key

   Expected result: A non-empty private .api-key file exists; existing key preserved. No secret is printed.

2. 2. Extract and verify the native chat template verbatim

   Decode the verified UTF-8 JSON, select its chat_template string and write those exact UTF-8 bytes without stripping whitespace or adding a newline. This is a documented derivation, not an independently hosted template. Compare the file back to the embedded string and its expected digest.

   CommandCopy command

   cd "$HOME/reka-flash3-hosting"
   python3 - <<'PYTEMPLATE'
   import json,pathlib,hashlib
   config=json.loads(pathlib.Path('models/tokenizer_config.json').read_text(encoding='utf-8'))
   text=config['chat_template']
   assert isinstance(text,str) and text
   raw=text.encode('utf-8')
   assert len(raw)==835 and hashlib.sha256(raw).hexdigest()=='1e330d18b71022b555c66089bebea6e4977d5fbc9a39534bb2e6e3d1af9829ed'
   pathlib.Path('models/reka-chat.jinja').write_bytes(raw)
   assert pathlib.Path('models/reka-chat.jinja').read_bytes()==config['chat_template'].encode('utf-8')
   print('Native template extracted verbatim and hash verified')
   PYTEMPLATE

   Expected result: Native template extracted verbatim and hash verified. The 835-byte template formats human:/assistant: rounds, joins a leading system message into the first user round, and strips completed prior-round reasoning traces.

Start the model

1. 1. Start one bounded native reasoning sequence

   Use the exact native template and raw content to retain <reasoning> tags. 8,192 total context includes prompt and output; reserve up to 4,096 generated tokens for both reasoning and final answer. This is a reduced smoke workload below the native 32,768 positions, not a full-capacity or successful-completion claim. No budget forcing or early </reasoning> injection is implemented.

   CommandCopy command

   cd "$HOME/reka-flash3-hosting"
   export REKA_FLASH3_SERVICE_KEY="$(cat .api-key)"
   ./llama.cpp/build/bin/llama-server -m models/reka-flash-3-Q4_K_M.gguf --alias reka-flash3 --host 127.0.0.1 --port 8080 --api-key "$REKA_FLASH3_SERVICE_KEY" --device none -ngl 0 -t 4 -c 8192 -np 1 --cache-type-k f16 --cache-type-v f16 --jinja --chat-template-file models/reka-chat.jinja --skip-chat-parsing --reasoning-format none --temp 0.6 --top-k 1024 --top-p 0.95 --min-p 0 --repeat-penalty 1 --frequency-penalty 0 --presence-penalty 0

   Expected result: Model initializes and loopback listener starts. Stop on tokenizer/architecture/template errors. The service key in process arguments may be visible to local/privileged users; use a trusted host.

Verify the setup

1. 1. Verify readiness, authenticated reasoning output and stop conditions

   Run in a second terminal on the server or its SSH shell. Explicit stop strings implement publisher guidance. Inspect raw choices[0].message.content for a completed </reasoning> block followed by a final answer of 36 counters; exact prose and token counts are not promised. A length finish, unclosed reasoning or absent/wrong final answer fails the smoke check. The cap does not guarantee reasoning completion.

   CommandCopy command

   cd "$HOME/reka-flash3-hosting"
   export REKA_FLASH3_SERVICE_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8080/health
   curl --fail http://127.0.0.1:8080/v1/chat/completions -H "Authorization: Bearer $REKA_FLASH3_SERVICE_KEY" -H 'Content-Type: application/json' --data-binary '{"model":"reka-flash3","messages":[{"role":"user","content":"A box contains 12 red counters, 18 blue counters and 6 green counters. How many counters are in the box? Think carefully and give a short final answer using only these numbers."}],"max_tokens":4096,"temperature":0.6,"top_k":1024,"top_p":0.95,"min_p":0,"repeat_penalty":1,"frequency_penalty":0,"presence_penalty":0,"stop":["<sep>","<|endoftext|>"]}' --output reasoning-response.json
   cat reasoning-response.json
   curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8080/v1/chat/completions -H 'Content-Type: application/json' --data-binary '{"model":"reka-flash3","messages":[{"role":"user","content":"Hello"}]}'

   Expected result: Ready health returns HTTP 200; authenticated generation returns nonempty raw content with completed native reasoning and a correct final total of 36, finish_reason not length. Missing service key returns HTTP 401. These are reader pass criteria, not observed inference results.

Performance evidence

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

Limits and troubleshooting

• Documented configuration only. No weights, installation/build, inference, fit, speed or quality observation was executed by the author.
• This is a dense Llama-compatible text model: 44 layers, hidden size 6,144, 64 attention heads, eight KV heads and head dimension 96. Native maximum positions 32,768 do not establish full-context fit in this 8K recipe.
• The publisher says Reka Flash thinks using <reasoning> tags and describes optional budget forcing by injecting a closing tag. This guide only caps total generated output at 4,096; it does not implement that technique or guarantee a final answer. Reasoning and final answer share the cap.
• Raw-content output preserves the native tags; do not expect a separate reasoning_content field. History formatting uses the pinned native template, which removes completed prior assistant reasoning traces.
• Stop on the literal <sep> string or <|endoftext|>; both are passed explicitly in the validation request. Do not replace native human:/assistant: formatting with a generic Llama instruction template.
• Native generation_config supplies temperature 0.6, top_k 1024 and top_p 0.95. Extra min-p/repetition/frequency/presence controls in this guide are author choices, not an additional publisher preset.
• The publisher recommends English use, warns about knowledge-intensive tasks and limited alignment/persona training. The supplied arithmetic task checks a narrow functional path, not general reliability, safety, coding or multilingual performance.
• The native/derivative metadata declare Apache-2.0. The Unsloth card names the Reka base without establishing an exact upstream conversion revision. Runnable derivative bytes and the independent publisher template are pinned separately.
• System packages resolve at installation. Preserve tool versions and build logs; GGML_NATIVE means build and run on the same CPU.
• 32 GiB RAM, 40 GiB disk and four threads are allocations, not measured minima or confirmed fit. Loopback/private bearer access assumes a trusted host; shared/public serving needs separate TLS, isolation and operations.
• Two guides do not establish complete model coverage. Other useful hardware/OS/runtime/full-context/budget-forcing/coding workflows remain unqualified, not excluded.

Checksum, tokenizer or template error

Stop. Confirm exact Unsloth file digest and native tokenizer JSON revision/hash. Re-extract chat_template verbatim and verify its derived SHA. Use the pinned native Llama runtime, not a generic prompt substitution.

Length finish, unclosed reasoning or no final answer

Treat the smoke check as failed. The reduced output budget is not guaranteed to complete reasoning. Record any changed output/context workload; do not silently cut tags or call a reasoning-only trace a final answer.

Memory pressure or context overflow

Stop and inspect host resources. Free/provision memory or explicitly reduce workload; file size does not prove fit. Preserve reasoning/output budget choices in any new configuration.

HTTP 401 or connection refused

Check foreground startup, loopback port 8080 and matching private service key. For cloud, verify the SSH tunnel points at the intended VM; keep public 8080 ingress closed.

Stop the model

1. 1. Stop the service

   Press Ctrl+C in server terminal. Unset key in each shell. Close the SSH tunnel with Ctrl+C. Stopping the process alone does not stop EC2 billing.

   CommandCopy command

   unset REKA_FLASH3_SERVICE_KEY

   Expected result: Server stops and localhost:8080 stops serving.

Uninstall

1. 1. Remove dedicated model, build and credential files

   Back up wanted logs/files first. Remove only this walkthrough directory after confirming its path. Shared apt packages can remain; do not remove other workloads.

   CommandCopy command

   rm -r -- "$HOME/reka-flash3-hosting"

   Expected result: Dedicated files and local service credential removed; shared system dependencies remain.

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/reka-flash-3