{
  "schema": "anvil-serving.qualification-summary/v1",
  "date": "2026-08-26",
  "repository_base_revision": "f0733d2e86f75e9f3c3b510c7d76c178659037e2",
  "model": "RadixArk/Qwen3.8-Flash-Next-NVFP4",
  "model_revision": "7b719225242aacd3dbd3f9407468c2ee9a9d2594",
  "image_digest": "sha256:59f06adce6f91401adf443bd168d45fdb2044d77671fd591c7c57a29d851cbae",
  "engine_revision": "d91c3682b0b429e4c70df63cd57f819588ce29b0",
  "sglang_qsa_patch_revision": "dac5523d1e5d2f4297fec40ef02fc76fb0f662d1",
  "hardware": "2x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition over PCIe without NVLink",
  "topology": "exclusive TP=2 under WSL2",
  "served_context_tokens": 262144,
  "client_prompt_tokens": 253952,
  "client_output_reserve_tokens": 8192,
  "concurrency": 1,
  "reasoning_contract": "thinking disabled",
  "kv_dtype": "auto (BF16 path)",
  "offload": "none",
  "current_recipe": "configs/qwen38-flash-next-radixark-nvfp4-sglang-sm120-qsa-fast-tp2-262k-mtp3-recipe.toml",
  "matched_control_recipe": "configs/qwen38-flash-next-radixark-nvfp4-sglang-sm120-qsa-fast-tp2-262k-nospec-recipe.toml",
  "performance": {
    "portable_qsa_no_spec": {
      "decode_tokens_per_second_4k_c1": 12.801,
      "decode_tokens_per_second_128k_c1": 12.544
    },
    "qsa_fast_no_spec": {
      "ttft_seconds_4k_c1": 0.14,
      "e2e_seconds_4k_c1": 0.69,
      "decode_tokens_per_second_4k_c1": 66.6,
      "ttft_seconds_128k_c1": 12.18,
      "prefill_tokens_per_second_128k_c1": 10057.0,
      "decode_tokens_per_second_128k_c1": 69.4,
      "full_reserve_prompt_tokens": 253703,
      "full_reserve_ttft_seconds": 27.131,
      "full_reserve_prefill_tokens_per_second": 9350.9,
      "full_reserve_decode_tokens_per_second": 66.9
    },
    "qsa_fast_mtp3": {
      "ttft_seconds_4k_c1": 0.15,
      "e2e_seconds_4k_c1": 0.36,
      "decode_tokens_per_second_4k_c1": 154.9,
      "ttft_seconds_128k_c1": 12.67,
      "prefill_tokens_per_second_128k_c1": 9664.0,
      "decode_tokens_per_second_128k_c1": 134.1,
      "full_reserve_prompt_tokens": 253703,
      "full_reserve_output_request_tokens": 8192,
      "full_reserve_ttft_seconds": 29.214,
      "full_reserve_prefill_tokens_per_second": 8684.3,
      "full_reserve_decode_tokens_per_second": 102.0
    },
    "mtp3_speedup_over_matched_control": {
      "decode_4k": 2.33,
      "decode_128k": 1.93
    },
    "mtp3_speedup_over_portable_qsa": {
      "decode_4k": 12.1,
      "decode_128k": 10.7
    }
  },
  "gates": {
    "exact_identity": true,
    "startup_patch_hash_and_import": true,
    "direct_functional": true,
    "long_context_retrieval_128k": true,
    "full_reserve_request": true,
    "tools": "20/20",
    "bounded_intelligence": "6/6",
    "session_continuation": "3/3",
    "repeated_tools": "3/3",
    "quality_regression_observed": false,
    "router_transition_identity": true,
    "hermes_client_acceptance": true,
    "pi_client_acceptance": true,
    "openclaw_client_acceptance": true,
    "openclaw_context_tokens": 262144,
    "pi_context_tokens": 262144
  },
  "failed_vllm_lane": {
    "model": "Inferact/Qwen3.8-Flash-Next-NVFP4",
    "revision": "103a7608316173ca6edd49929544244de7ffda70",
    "attempts": [
      "default NCCL initialization failed before weight load",
      "WSL2 NCCL controls exposed V2 runner UVA failure before weight load",
      "V1 runner loaded approximately 85.76 GiB per card, then compile autotuning OOMed before KV allocation"
    ],
    "decision": "exact tested vLLM recipes empirically disqualified; no universal checkpoint or 262K infeasibility claim"
  },
  "decision": "human-authorized current text Primary on the locally qualified QSA-fast MTP3 recipe; no multimodal promotion"
}