models.yaml

schema_version: 3

# These curated entries preserve known profiles, reasoning policies, and
# operator-selected contexts. They take precedence over automatic runtime discovery, but are not
# an allowlist: any other installed Ollama tag can be resolved read-only at run
# time when its effective num_ctx is present in runtime metadata.
models:
  laguna-apex-128k:
    display_name: Laguna-S-2.1-Uncensored APEX
    provider: gx10
    runtime: ollama
    runtime_model: coder-uncens:latest
    runtime_digest: null
    quantization: existing-local
    context_length: 131072
    reasoning_effort: high
    reasoning_policy: effort:high
    supported_reasoning_policies: ["off", native, effort:high]
    role: example-quality-profile
    notes:
      - A null runtime_digest resolves and records the installed digest at run time.

  agent-main:latest:
    display_name: Qwen3.6-35B-A3B — 36.0B — Q8_0 — 65536-token context
    provider: gx10
    runtime: ollama
    runtime_model: agent-main:latest
    runtime_digest: sha256:0218f872e86baa9c7610509f27db36a7bc52eea7afee24688f81ca74ffcb6c77
    source_model: Qwen3.6-35B-A3B
    source_version: "3.6"
    architecture: qwen35moe
    parameter_variant: 36.0B
    parameter_count: 35951822704
    quantization: Q8_0
    context_length: 65536
    native_context_length: 262144
    reasoning_effort: medium
    reasoning_policy: effort:medium
    supported_reasoning_policies: ["off", native, effort:medium]
    role: curated-runtime-profile
    notes:
      - Context is operator-configured; the runtime reports a 262144-token native maximum and no stored num_ctx.

  coder-max:latest:
    display_name: Qwen3-Coder-Next — 79.7B — Q8_0 — 65536-token context
    provider: gx10
    runtime: ollama
    runtime_model: coder-max:latest
    runtime_digest: sha256:3f68e12b44eea7c1f464501436bbbe67a8234bf2694efc1d7407ab3df73251b6
    source_model: qwen3-coder-next:q8_0
    source_version: Next
    architecture: qwen3next
    parameter_variant: 79.7B
    parameter_count: 79674391296
    quantization: Q8_0
    context_length: 65536
    native_context_length: 262144
    reasoning_effort: medium
    reasoning_policy: effort:medium
    supported_reasoning_policies: ["off", native, effort:medium]
    role: curated-runtime-profile
    notes:
      - Context is operator-configured; the runtime reports a 262144-token native maximum and no stored num_ctx.

  coder-max-128k:latest:
    display_name: Qwen3-Coder-Next — 79.7B — Q8_0 — 131072-token context
    provider: gx10
    runtime: ollama
    runtime_model: coder-max-128k:latest
    runtime_digest: sha256:0dbde7ee79ca3ede12b17811e23a4d18e9e61a5dfa261e831cd3f6cdeeaf1803
    source_model: qwen3-coder-next:q8_0
    source_version: Next
    architecture: qwen3next
    parameter_variant: 79.7B
    parameter_count: 79674391296
    quantization: Q8_0
    context_length: 131072
    native_context_length: 262144
    reasoning_effort: medium
    reasoning_policy: effort:medium
    supported_reasoning_policies: ["off", native, effort:medium]
    role: curated-runtime-profile

  reviewer-deep:latest:
    display_name: GPT-OSS-120B — 116.8B — MXFP4 — 65536-token context
    provider: gx10
    runtime: ollama
    runtime_model: reviewer-deep:latest
    runtime_digest: sha256:a951a23b46a1f6093dafee2ea481d634b4e31ac720a8a16f3f91e04f5a40ecd9
    source_model: GPT-OSS-120B
    source_version: 120B
    architecture: gptoss
    parameter_variant: 116.8B
    parameter_count: 116829156672
    quantization: MXFP4
    context_length: 65536
    native_context_length: 131072
    reasoning_effort: medium
    reasoning_policy: effort:medium
    supported_reasoning_policies: [native, effort:low, effort:medium, effort:high]
    role: curated-runtime-profile
    notes:
      - Context is operator-configured; the runtime reports a 131072-token native maximum and no stored num_ctx.

  coder-uncens:latest:
    display_name: Laguna-S-2.1-Uncensored APEX — 117.6B — Q6_K — 131072-token context
    provider: gx10
    runtime: ollama
    runtime_model: coder-uncens:latest
    runtime_digest: sha256:50197a047af0f7c198929eb80bbd4bb5a44cd43ca878f5241c846ffbf3ee9a7b
    source_model: Laguna-S-2.1-Uncensored APEX
    source_version: "2.1"
    architecture: laguna
    parameter_variant: 117.6B
    parameter_count: 117561977600
    quantization: Q6_K
    context_length: 131072
    native_context_length: 1048576
    reasoning_effort: high
    reasoning_policy: effort:high
    supported_reasoning_policies: ["off", native, effort:high]
    role: curated-runtime-profile

  coder-uncens-256k:latest:
    display_name: Laguna-S-2.1-Uncensored APEX — 117.6B — Q6_K — 262144-token context
    provider: gx10
    runtime: ollama
    runtime_model: coder-uncens-256k:latest
    runtime_digest: sha256:ea34631ee5c00f9acd0182cb9503098e96c551e3841bf537bbc69a4576aea84b
    source_model: Laguna-S-2.1-Uncensored APEX
    source_version: "2.1"
    architecture: laguna
    parameter_variant: 117.6B
    parameter_count: 117561977600
    quantization: Q6_K
    context_length: 262144
    native_context_length: 1048576
    reasoning_effort: high
    reasoning_policy: effort:high
    supported_reasoning_policies: ["off", native, effort:high]
    role: curated-runtime-profile

  coder-uncens-qwen:latest:
    display_name: Huihui Qwen3-Coder-Next Abliterated — 79.7B — Q8_0 — 131072-token context
    provider: gx10
    runtime: ollama
    runtime_model: coder-uncens-qwen:latest
    runtime_digest: sha256:91e5aee4586a37fda420039bfc30293371bab72026a06239e1183339caff2ae0
    source_model: huihui_ai/qwen3-coder-next-abliterated:q8_0
    source_version: Next
    architecture: qwen3next
    parameter_variant: 79.7B
    parameter_count: 79674391296
    quantization: Q8_0
    context_length: 131072
    native_context_length: 262144
    reasoning_effort: medium
    reasoning_policy: effort:medium
    supported_reasoning_policies: ["off", native, effort:medium]
    role: curated-runtime-profile

  hermes4-70b:latest:
    display_name: Hermes-4-70B — 70.6B — Q4_K_M — 65536-token context
    provider: gx10
    runtime: ollama
    runtime_model: hermes4-70b:latest
    runtime_digest: sha256:59c3de79a9d1198dd7065fe1431761274b0bd4fe3582c0c8b59ec044be9627ea
    source_model: Hermes-4-70B
    source_version: "4"
    architecture: llama
    parameter_variant: 70.6B
    parameter_count: 70553706560
    quantization: Q4_K_M
    context_length: 65536
    native_context_length: 131072
    reasoning_effort: medium
    reasoning_policy: effort:medium
    supported_reasoning_policies: ["off", native, effort:medium]
    role: curated-runtime-profile
    notes:
      - Context is operator-configured; the runtime reports a 131072-token native maximum and no stored num_ctx.

  qwen38-q8-medium-262k:
    display_name: Qwen3.8 0814 — 27.3B — Q8_0 — 262144-token context
    provider: gx10
    runtime: ollama
    runtime_model: qwen38-q8-262k:latest
    runtime_digest: sha256:4ab95509a27d7a3f23dcc612a660858e9f28c1a5322bd9240f34559bcf888988
    source_model: qwen3.8:27b-q8_0
    source_version: "0814"
    architecture: qwen35
    parameter_variant: 27.3B
    quantization: Q8_0
    context_length: 262144
    reasoning_effort: medium
    reasoning_policy: effort:medium
    supported_reasoning_policies: ["off", native, effort:medium]
    role: reproduced-standard-profile
    default_for_runtime: true
    notes:
      - The immutable digest reproduces the existing completed standard result.

  qwen38-q8-medium-128k:
    display_name: Qwen3.8 0814 — 27.3B — Q8_0 — 131072-token context
    provider: gx10
    runtime: ollama
    runtime_model: qwen38-q8-262k:latest
    runtime_digest: sha256:4ab95509a27d7a3f23dcc612a660858e9f28c1a5322bd9240f34559bcf888988
    source_model: qwen3.8:27b-q8_0
    source_version: "0814"
    architecture: qwen35
    parameter_variant: 27.3B
    quantization: Q8_0
    context_length: 131072
    reasoning_effort: medium
    reasoning_policy: effort:medium
    supported_reasoning_policies: ["off", native, effort:medium]
    role: reproduced-128k-profile
    notes:
      - This profile shares weights with the 262K entry but freezes a 128K context.

  gemma4:31b-it-bf16:
    display_name: Gemma 4 31B IT — 31.3B — F16 — 262144-token context
    provider: gx10
    runtime: ollama
    runtime_model: gemma4:31b-it-bf16
    runtime_digest: sha256:236d76ae08745dbc143c31b9271b0f25750885199aa6039d0fc0113171606e6d
    source_model: gemma4:31b-it-bf16
    source_version: "4"
    architecture: gemma4
    parameter_variant: 31.3B
    parameter_count: 31273089132
    quantization: F16
    context_length: 262144
    native_context_length: 262144
    template_sha256: sha256:b507b9c2f6ca642bffcd06665ea7c91f235fd32daeefdf875a0f938db05fb315
    runtime_capabilities: [completion, thinking, tools, vision]
    reasoning_effort: native
    reasoning_policy: native
    supported_reasoning_policies: ["off", native]
    role: curated-runtime-profile
    notes:
      - Native and off behavior were verified through bounded OpenAI-compatible probes on 2026-08-28; no effort level is inferred.
      - The configured smoke on 2026-08-28 preserved an infrastructure diagnostic after both BenchLocal requests exceeded the fixed 180-second case limit; this deployment is currently impractical for GX10 agentic use.

  ornith15-q8:latest:
    display_name: Ornith-1.5-35B-A3B — 35.5B — Q8_0 — 262144-token context
    provider: gx10
    runtime: ollama
    runtime_model: ornith15-q8:latest
    runtime_digest: sha256:c7c57f189918400a4b4193530cceec9136f95c019357de02abc61018f8bad485
    source_model: ornith-ai/Ornith-1.5-35B-A3B
    source_version: "1.5"
    architecture: qwen35moe
    parameter_variant: 35.5B
    parameter_count: 35505251456
    quantization: Q8_0
    context_length: 262144
    native_context_length: 262144
    template_sha256: sha256:f55f52930aa8bf44ab5cb85f99370fcc3c56e9a85640b812086d5330bce5d86b
    runtime_capabilities: [completion, thinking, tools, vision]
    reasoning_effort: none
    reasoning_policy: "off"
    supported_reasoning_policies: ["off", native]
    role: curated-runtime-profile
    notes:
      - Native and off behavior were verified through bounded OpenAI-compatible probes on 2026-08-28; no effort level is inferred.
      - The intended Hermes deployment disables reasoning; native remains available only as an explicit diagnostic policy.

  deepseek-v4-flash:
    display_name: DeepSeek V4 Flash — DS4 GGUF + DSpark — 524288-token context
    provider: gx10
    runtime: ds4
    runtime_version: "0.6.5"
    runtime_model: deepseek-v4-flash
    runtime_digest: sha256:ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0
    source_model: DeepSeek-V4-Flash
    source_version: V4-Flash
    architecture: deepseek
    parameter_variant: DeepSeek-V4-Flash
    quantization: DS4-GGUF
    context_length: 524288
    reasoning_effort: none
    reasoning_policy: "off"
    supported_reasoning_policies: ["off"]
    endpoint_port: 8000
    api_mode: openai-chat-completions
    capabilities:
      streaming: true
      tool_calls: true
    deployment_artifacts:
      base_gguf_sha256: sha256:ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0
      dspark_drafter_sha256: sha256:8fa269560dc76fd73e4233ad9b1938b5f65dd363381fd9b1a5c6183f7d12d686
      dspark_enabled: true
    role: curated-runtime-profile
    notes:
      - DS4 exposes only /v1/models and /v1/chat/completions to the harness.
      - Thinking-disabled control serializes as reasoning_effort=none.

  qwen38-flash-next-nvfp4-262k:
    display_name: Qwen3.8-Flash-Next — 125B main + 51B PLE / 6B active — NVFP4 — 262144-token context
    provider: gx10
    runtime: vllm
    runtime_version: blazux/qwen3.8-Flash-DGX@209646c
    runtime_model: qwen3.8-flash-next
    runtime_digest: sha256:56d617b9008714d82b491a73c3fc8f86501a2a4dcd17f0b883359762f1af3e17
    runtime_digest_kind: vllm-deployment-provenance
    canonical_name: Qwen3.8-Flash-Next
    source_model: RadixArk/Qwen3.8-Flash-Next-NVFP4
    source_version: Flash-Next
    architecture: Qwen3.8-Flash-Next sparse MoE
    parameter_variant: 125B main + 51B PLE / 6B active
    quantization: NVFP4
    context_length: 262144
    native_context_length: 262144
    reasoning_effort: native
    reasoning_policy: native
    supported_reasoning_policies: ["off", native]
    reasoning_control_profile: qwen-enable-thinking-v1
    endpoint_port: 18300
    api_mode: openai-chat-completions
    capabilities:
      streaming: true
      tool_calls: true
    serving_configuration:
      gpu_memory_utilization: 0.75
      max_num_seqs: 2
      mtp_speculative_tokens: 2
      prefix_caching: true
      exact_topk: true
      prewarm: false
    deployment_artifacts:
      provenance_contract: vllm-deployment-provenance-v1
      checkpoint: RadixArk/Qwen3.8-Flash-Next-NVFP4
      checkpoint_revision: 7b719225242aacd3dbd3f9407468c2ee9a9d2594
      serving_implementation: blazux/qwen3.8-Flash-DGX
      serving_revision: 209646c
      serving_revision_full: 209646cd98290035ccbddef29b14c460460a8709
    role: curated-runtime-profile
    default_for_runtime: false
    notes:
      - vLLM exposes only /v1/models and /v1/chat/completions to the harness.
      - runtime_digest is the SHA-256 of the versioned immutable deployment-provenance descriptor, not a weight-file checksum.
      - serving_configuration records the live deployment settings but is intentionally excluded from the immutable provenance digest.
      - Native reasoning uses the checkpoint template default; off uses chat_template_kwargs.enable_thinking=false.
      - Historical native-thinking deployment with MTP/speculative decoding set to 2; retained for exact attribution of v4 results.

  qwen38-flash-next-nvfp4-262k-quality:
    display_name: Qwen3.8-Flash-Next — quality deployment (thinking off, MTP off) — NVFP4 — 262144-token context
    provider: gx10
    runtime: vllm
    runtime_version: blazux/qwen3.8-Flash-DGX@209646c
    runtime_model: qwen3.8-flash-next
    runtime_digest: sha256:56d617b9008714d82b491a73c3fc8f86501a2a4dcd17f0b883359762f1af3e17
    runtime_digest_kind: vllm-deployment-provenance
    canonical_name: Qwen3.8-Flash-Next
    source_model: RadixArk/Qwen3.8-Flash-Next-NVFP4
    source_version: Flash-Next
    architecture: Qwen3.8-Flash-Next sparse MoE
    parameter_variant: 125B main + 51B PLE / 6B active
    quantization: NVFP4
    context_length: 262144
    native_context_length: 262144
    reasoning_effort: none
    reasoning_policy: "off"
    supported_reasoning_policies: ["off", native]
    reasoning_control_profile: qwen-enable-thinking-v1
    endpoint_port: 18300
    api_mode: openai-chat-completions
    capabilities:
      streaming: true
      tool_calls: true
    serving_configuration:
      gpu_memory_utilization: 0.75
      max_num_seqs: 2
      mtp_speculative_tokens: 0
      prefix_caching: true
      exact_topk: true
      prewarm: false
      kv_cache_dtype: BF16/default
    deployment_artifacts:
      provenance_contract: vllm-deployment-provenance-v1
      checkpoint: RadixArk/Qwen3.8-Flash-Next-NVFP4
      checkpoint_revision: 7b719225242aacd3dbd3f9407468c2ee9a9d2594
      serving_implementation: blazux/qwen3.8-Flash-DGX
      serving_revision: 209646c
      serving_revision_full: 209646cd98290035ccbddef29b14c460460a8709
    role: curated-runtime-profile
    default_for_runtime: true
    notes:
      - Current quality-oriented deployment identity for gx10-qualification-v5.
      - The immutable provenance digest is shared with the historical alias because checkpoint, quantization, runtime model ID, and serving implementation revision are unchanged.
      - Mutable deployment provenance separately records thinking off and MTP/speculative decoding off.
      - Thinking off serializes as chat_template_kwargs.enable_thinking=false through the vLLM/Qwen reasoning adapter.