Hermes 4 70B

Exact model and requirements

Artifact:

Unsloth (GGUF derivative); Nous Research (native model) / Hermes 4 70B Q4_K_M, Unsloth derivative · checkpoint unsloth/Hermes-4-70B-GGUF

Publisher revision:

b5483741203956fbc0557a26429ee73bf4cb7944

Artifact digest:

sha256:1bdf4b8469bfa36b274412818e47bf1a39d667d1d9d9558bcdb3ecae2bc9fedc

Download size:

39.6 GiB

Memory requirement (unknown):

Unknown — No exact CPU peak for this dense 70B Q4_K_M artifact and 4K 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:

4,096 token context; Unknown; Unknown; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Embedded Llama-family GGUF tokenizer with explicit pinned Nous Research native template · GGUF b5483741203956fbc0557a26429ee73bf4cb7944; native template d5dec2bd6b3930a09ddefd0b7fc6523fe0720d09

Sampling / reasoning:

Publisher card preset: temperature 0.6, top_p 0.95, top_k 20. Native generation_config instead lists top_p 0.9; this guide explicitly chooses the card preset. min_p 0, repeat_penalty 1 and frequency/presence penalties 0 are author controls. Native template thinking=false selects the standard system prompt, not a guaranteed no-reasoning model mode. Raw output preserves any <think> tags. 512 output tokens can truncate deliberation or final content; a complete correct answer is required.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Both native and Unsloth metadata declare llama3, while the native card identifies Llama 3.1 as the base. The native root LICENSE is absent (404). Retain this discrepancy: this guide does not replace the publisher label or claim the two licenses equivalent. Read and comply with the Meta Llama 3 and Llama 3.1 Community License Agreements and both incorporated acceptable-use policies, including applicable attribution/notice, Built with branding, redistribution/model-naming, prohibited-use, other-model training restrictions and 700-million monthly-active-user commercial-license conditions. Both provide conditional use grants; this is restricted open weights, not Apache or unrestricted open source. Stop if your use cannot satisfy the applicable terms. No legal-clearance claim is made.

• Nous Research native chat_template.jinja · 0 GiB · revision d5dec2bd6b3930a09ddefd0b7fc6523fe0720d09

• The 42,520,399,008-byte weights do not establish working memory. Cache, compute buffers, runtime and OS need additional resources.
• 64 GiB RAM, 80 GiB disk and eight CPU threads are planning allocations, not measured minima or confirmed fit/performance.

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 "$HOME/hermes4-70b-hosting/models"
   cd "$HOME/hermes4-70b-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 and verify the exact derivative and native template

   No authentication is required for these public artifacts. Stop on either digest mismatch. The Unsloth card advertises template fixes; this guide explicitly overrides its embedded template with the independently pinned native Nous Research file. Exact conversion provenance is not inferred.

   CommandCopy command

   cd "$HOME/hermes4-70b-hosting"
   curl --fail --location --output models/Hermes-4-70B-Q4_K_M.gguf 'https://huggingface.co/unsloth/Hermes-4-70B-GGUF/resolve/b5483741203956fbc0557a26429ee73bf4cb7944/Hermes-4-70B-Q4_K_M.gguf'
   curl --fail --location --output models/hermes-chat.jinja 'https://huggingface.co/NousResearch/Hermes-4-70B/raw/d5dec2bd6b3930a09ddefd0b7fc6523fe0720d09/chat_template.jinja'
   sha256sum --check <<'SHA256'
   1bdf4b8469bfa36b274412818e47bf1a39d667d1d9d9558bcdb3ecae2bc9fedc  models/Hermes-4-70B-Q4_K_M.gguf
   89116c1ac22ce005afdf1772cb40644c1529af6a91300cf655566d10274d4175  models/hermes-chat.jinja
   SHA256

   Expected result: Both files report OK. GGUF is 42,520,399,008 bytes; native template is 4,121 bytes. No projector is needed for this text configuration.

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/hermes4-70b-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.

Start the model

1. 1. Start bounded standard-template chat

   The native thinking=false branch uses the standard Hermes system prompt. Keep raw response content and do not promise absence of reasoning. Context is 4,096 total tokens shared by prompt and output, below the native 131,072 positional limit. Leave this process in the foreground.

   CommandCopy command

   cd "$HOME/hermes4-70b-hosting"
   export HERMES4_SERVICE_KEY="$(cat .api-key)"
   ./llama.cpp/build/bin/llama-server -m models/Hermes-4-70B-Q4_K_M.gguf --alias hermes4-70b --host 127.0.0.1 --port 8080 --api-key "$HERMES4_SERVICE_KEY" --device none -ngl 0 -t 8 -c 4096 -np 1 --cache-type-k f16 --cache-type-v f16 --jinja --chat-template-file models/hermes-chat.jinja --chat-template-kwargs '{"thinking":false,"keep_cots":false}' --skip-chat-parsing --reasoning-format none --temp 0.6 --top-k 20 --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 can be visible to trusted local or privileged users.

Verify the setup

1. 1. Check readiness, authenticated chat and missing-key rejection

   Run in a second terminal on the same desktop or VM SSH shell. Inspect raw choices[0].message.content: a complete final answer must give 36 counters. Exact generated prose and reasoning are unknown; a length finish or unclosed reasoning trace fails this reduced smoke check.

   CommandCopy command

   cd "$HOME/hermes4-70b-hosting"
   export HERMES4_SERVICE_KEY="$(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 $HERMES4_SERVICE_KEY" -H 'Content-Type: application/json' --data-binary '{"model":"hermes4-70b","messages":[{"role":"user","content":"A box has 12 red counters, 18 blue counters and 6 green counters. Give the total number of counters in one short sentence."}],"max_tokens":512,"temperature":0.6,"top_k":20,"top_p":0.95,"min_p":0,"repeat_penalty":1,"frequency_penalty":0,"presence_penalty":0,"stop":["<|eot_id|>","<|end_of_text|>"]}' --output chat-response.json
   cat chat-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":"hermes4-70b","messages":[{"role":"user","content":"Hello"}]}'

   Expected result: Ready health HTTP 200, authenticated nonempty complete answer 36 counters with finish_reason not length, and missing-key HTTP 401. These are reader pass criteria, not observed 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 measurement was performed.
• Dense Llama architecture: 80 layers, hidden size 8,192, 64 attention heads, eight KV heads, head dimension 128 and Llama 3.1 rope scaling. Full 70B weights load; no active-parameter memory discount.
• The exact native template is passed separately. Unsloth advertises template fixes, but their embedded template is not assumed identical to the native file. Both sources are pinned independently.
• thinking=false selects a standard system prompt. Hermes remains hybrid reasoning-capable; any reasoning shares the 512-token cap with the final answer. Raw output preserves <think> tags; no separate reasoning_content is promised.
• Native template system messages can override its standard/thinking prompt. This single-turn request supplies no system message or tools. Native thinking=true, preserved-chain history, tool calling and structured JSON are useful separate workflows that remain unqualified here.
• The model card recommends temperature 0.6, top_p 0.95 and top_k 20. The generation config instead says top_p 0.9; this guide explicitly uses the card preset.
• Both native and Unsloth metadata declare llama3, while the native card identifies Llama 3.1 as the base. The native root LICENSE is absent (404). Retain this discrepancy: this guide does not replace the publisher label or claim the two licenses equivalent. Read and comply with the Meta Llama 3 and Llama 3.1 Community License Agreements and both incorporated acceptable-use policies, including applicable attribution/notice, Built with branding, redistribution/model-naming, prohibited-use, other-model training restrictions and 700-million monthly-active-user commercial-license conditions. Both provide conditional use grants; this is restricted open weights, not Apache or unrestricted open source. Stop if your use cannot satisfy the applicable terms. No legal-clearance claim is made.
• The Unsloth derivative names Hermes 4 but does not establish an exact native conversion revision. The binary digest and independent native template pin are authoritative for this recipe.
• System packages resolve at installation; retain versions/build logs. GGML_NATIVE requires building on the CPU that will run the binary.
• 64 GiB RAM, 80 GiB disk and eight threads are planning allocations, not measured minima or fit guarantees. Loopback/bearer/SSH access assumes a trusted host; public multiuser serving needs separate operations.
• These two guides do not establish exhaustive model coverage. Full context, reasoning, tools, other runtimes and accelerators remain unqualified.

Checksum or template error

Stop and verify both independent revisions and hashes. Use the pinned native template with --jinja; do not substitute generic chat formatting.

Truncated answer or unclosed reasoning

Treat the smoke check as failed. Standard-template mode is not a no-reasoning guarantee. A larger output/context allocation is a distinct workload and must be recorded.

Memory pressure

Stop; inspect actual resources. File size is not working RAM. Increase resources or explicitly qualify a different workload/quantization; do not infer fit.

HTTP 401 or connection refused

Check foreground initialization and matching private key. For cloud, check the SSH tunnel to 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 HERMES4_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/hermes4-70b-hosting"

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

Related guides

• https://opentools.ai/llms/hermes-4-70b

Exact model and requirements

Artifact:

Unsloth (GGUF derivative); Nous Research (native model) / Hermes 4 70B Q4_K_M, Unsloth derivative · checkpoint unsloth/Hermes-4-70B-GGUF

Publisher revision:

b5483741203956fbc0557a26429ee73bf4cb7944

Artifact digest:

sha256:1bdf4b8469bfa36b274412818e47bf1a39d667d1d9d9558bcdb3ecae2bc9fedc

Download size:

39.6 GiB

Memory requirement (unknown):

Unknown — No exact CPU peak for this dense 70B Q4_K_M artifact and 4K 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:

4,096 token context; Unknown; Unknown; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Embedded Llama-family GGUF tokenizer with explicit pinned Nous Research native template · GGUF b5483741203956fbc0557a26429ee73bf4cb7944; native template d5dec2bd6b3930a09ddefd0b7fc6523fe0720d09

Sampling / reasoning:

Publisher card preset: temperature 0.6, top_p 0.95, top_k 20. Native generation_config instead lists top_p 0.9; this guide explicitly chooses the card preset. min_p 0, repeat_penalty 1 and frequency/presence penalties 0 are author controls. Native template thinking=false selects the standard system prompt, not a guaranteed no-reasoning model mode. Raw output preserves any <think> tags. 512 output tokens can truncate deliberation or final content; a complete correct answer is required.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Both native and Unsloth metadata declare llama3, while the native card identifies Llama 3.1 as the base. The native root LICENSE is absent (404). Retain this discrepancy: this guide does not replace the publisher label or claim the two licenses equivalent. Read and comply with the Meta Llama 3 and Llama 3.1 Community License Agreements and both incorporated acceptable-use policies, including applicable attribution/notice, Built with branding, redistribution/model-naming, prohibited-use, other-model training restrictions and 700-million monthly-active-user commercial-license conditions. Both provide conditional use grants; this is restricted open weights, not Apache or unrestricted open source. Stop if your use cannot satisfy the applicable terms. No legal-clearance claim is made.

• Nous Research native chat_template.jinja · 0 GiB · revision d5dec2bd6b3930a09ddefd0b7fc6523fe0720d09

• The 42,520,399,008-byte weights do not establish working memory. Cache, compute buffers, runtime and OS need additional resources.
• 64 GiB RAM, 80 GiB disk and eight CPU threads are planning allocations, not measured minima or confirmed fit/performance.

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.2xlarge 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 "$HOME/hermes4-70b-hosting/models"
   cd "$HOME/hermes4-70b-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 and verify the exact derivative and native template

   No authentication is required for these public artifacts. Stop on either digest mismatch. The Unsloth card advertises template fixes; this guide explicitly overrides its embedded template with the independently pinned native Nous Research file. Exact conversion provenance is not inferred.

   CommandCopy command

   cd "$HOME/hermes4-70b-hosting"
   curl --fail --location --output models/Hermes-4-70B-Q4_K_M.gguf 'https://huggingface.co/unsloth/Hermes-4-70B-GGUF/resolve/b5483741203956fbc0557a26429ee73bf4cb7944/Hermes-4-70B-Q4_K_M.gguf'
   curl --fail --location --output models/hermes-chat.jinja 'https://huggingface.co/NousResearch/Hermes-4-70B/raw/d5dec2bd6b3930a09ddefd0b7fc6523fe0720d09/chat_template.jinja'
   sha256sum --check <<'SHA256'
   1bdf4b8469bfa36b274412818e47bf1a39d667d1d9d9558bcdb3ecae2bc9fedc  models/Hermes-4-70B-Q4_K_M.gguf
   89116c1ac22ce005afdf1772cb40644c1529af6a91300cf655566d10274d4175  models/hermes-chat.jinja
   SHA256

   Expected result: Both files report OK. GGUF is 42,520,399,008 bytes; native template is 4,121 bytes. No projector is needed for this text configuration.

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/hermes4-70b-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.

Start the model

1. 1. Start bounded standard-template chat

   The native thinking=false branch uses the standard Hermes system prompt. Keep raw response content and do not promise absence of reasoning. Context is 4,096 total tokens shared by prompt and output, below the native 131,072 positional limit. Leave this process in the foreground.

   CommandCopy command

   cd "$HOME/hermes4-70b-hosting"
   export HERMES4_SERVICE_KEY="$(cat .api-key)"
   ./llama.cpp/build/bin/llama-server -m models/Hermes-4-70B-Q4_K_M.gguf --alias hermes4-70b --host 127.0.0.1 --port 8080 --api-key "$HERMES4_SERVICE_KEY" --device none -ngl 0 -t 8 -c 4096 -np 1 --cache-type-k f16 --cache-type-v f16 --jinja --chat-template-file models/hermes-chat.jinja --chat-template-kwargs '{"thinking":false,"keep_cots":false}' --skip-chat-parsing --reasoning-format none --temp 0.6 --top-k 20 --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 can be visible to trusted local or privileged users.

Verify the setup

1. 1. Check readiness, authenticated chat and missing-key rejection

   Run in a second terminal on the same desktop or VM SSH shell. Inspect raw choices[0].message.content: a complete final answer must give 36 counters. Exact generated prose and reasoning are unknown; a length finish or unclosed reasoning trace fails this reduced smoke check.

   CommandCopy command

   cd "$HOME/hermes4-70b-hosting"
   export HERMES4_SERVICE_KEY="$(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 $HERMES4_SERVICE_KEY" -H 'Content-Type: application/json' --data-binary '{"model":"hermes4-70b","messages":[{"role":"user","content":"A box has 12 red counters, 18 blue counters and 6 green counters. Give the total number of counters in one short sentence."}],"max_tokens":512,"temperature":0.6,"top_k":20,"top_p":0.95,"min_p":0,"repeat_penalty":1,"frequency_penalty":0,"presence_penalty":0,"stop":["<|eot_id|>","<|end_of_text|>"]}' --output chat-response.json
   cat chat-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":"hermes4-70b","messages":[{"role":"user","content":"Hello"}]}'

   Expected result: Ready health HTTP 200, authenticated nonempty complete answer 36 counters with finish_reason not length, and missing-key HTTP 401. These are reader pass criteria, not observed 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 measurement was performed.
• Dense Llama architecture: 80 layers, hidden size 8,192, 64 attention heads, eight KV heads, head dimension 128 and Llama 3.1 rope scaling. Full 70B weights load; no active-parameter memory discount.
• The exact native template is passed separately. Unsloth advertises template fixes, but their embedded template is not assumed identical to the native file. Both sources are pinned independently.
• thinking=false selects a standard system prompt. Hermes remains hybrid reasoning-capable; any reasoning shares the 512-token cap with the final answer. Raw output preserves <think> tags; no separate reasoning_content is promised.
• Native template system messages can override its standard/thinking prompt. This single-turn request supplies no system message or tools. Native thinking=true, preserved-chain history, tool calling and structured JSON are useful separate workflows that remain unqualified here.
• The model card recommends temperature 0.6, top_p 0.95 and top_k 20. The generation config instead says top_p 0.9; this guide explicitly uses the card preset.
• Both native and Unsloth metadata declare llama3, while the native card identifies Llama 3.1 as the base. The native root LICENSE is absent (404). Retain this discrepancy: this guide does not replace the publisher label or claim the two licenses equivalent. Read and comply with the Meta Llama 3 and Llama 3.1 Community License Agreements and both incorporated acceptable-use policies, including applicable attribution/notice, Built with branding, redistribution/model-naming, prohibited-use, other-model training restrictions and 700-million monthly-active-user commercial-license conditions. Both provide conditional use grants; this is restricted open weights, not Apache or unrestricted open source. Stop if your use cannot satisfy the applicable terms. No legal-clearance claim is made.
• The Unsloth derivative names Hermes 4 but does not establish an exact native conversion revision. The binary digest and independent native template pin are authoritative for this recipe.
• System packages resolve at installation; retain versions/build logs. GGML_NATIVE requires building on the CPU that will run the binary.
• 64 GiB RAM, 80 GiB disk and eight threads are planning allocations, not measured minima or fit guarantees. Loopback/bearer/SSH access assumes a trusted host; public multiuser serving needs separate operations.
• These two guides do not establish exhaustive model coverage. Full context, reasoning, tools, other runtimes and accelerators remain unqualified.

Checksum or template error

Stop and verify both independent revisions and hashes. Use the pinned native template with --jinja; do not substitute generic chat formatting.

Truncated answer or unclosed reasoning

Treat the smoke check as failed. Standard-template mode is not a no-reasoning guarantee. A larger output/context allocation is a distinct workload and must be recorded.

Memory pressure

Stop; inspect actual resources. File size is not working RAM. Increase resources or explicitly qualify a different workload/quantization; do not infer fit.

HTTP 401 or connection refused

Check foreground initialization and matching private key. For cloud, check the SSH tunnel to 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 HERMES4_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/hermes4-70b-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/hermes-4-70b