{
  "provenance": {
    "kind": "transcribed deployment observations, not per-request raw benchmark data",
    "sources": ["Thor thorllmtest/README.md inspected 2026-11-10", "operator test notes supplied 2026-11-10"],
    "test_dates": ["2026-11-04", "2026-11-05"],
    "limitations": ["few requests per configuration", "different vLLM versions", "prefix caching enabled; cache states not isolated", "no controlled accuracy evaluation", "no MTP performance measurements"]
  },
  "environment": {"device": "Jetson AGX Thor", "marketed_unified_memory_gb": 128, "sm": "110", "l4t": "R39.2", "host_cuda_driver_api": "13.2", "power_mode": "MAXN", "context_limit_tokens": 65536, "max_active_sequences": 2, "kv_dtype": "fp8", "mode": "language-model-only"},
  "qwen35": {
    "model_id": "nvidia/Qwen3.5-122B-A10B-NVFP4",
    "revision": "98915d837c4e7c87ac8296d02e89de19b3207e6d",
    "image": "vllm/vllm-openai:v0.28.0",
    "weights_gib": 77.79,
    "runtime_model_gib": 72.37,
    "kv_cache_gib_64k_config": 12.02,
    "kv_cache_token_capacity_reported": 747706,
    "single_stream_approx_tok_s": 16,
    "two_stream_aggregate_tok_s_observed_range": [27.79, 31.9],
    "short_input_ttft_s": 0.403,
    "long_input_ttft_s": {"approx_14k": 7.6, "approx_30k": 19.3, "approx_60k": 50.5},
    "one_moe_explanation_example": {"tokens": 2034, "elapsed_s": 125}
  },
  "flash_next": {
    "model_id": "nvidia/Qwen3.8-Flash-Next-NVFP4",
    "revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47",
    "image": "vllm/vllm-openai@sha256:cca10003856866925653fb4ec45309df385dc15f802dbd47c289eb9c5720f197",
    "vllm_version": "0.33.0rc1.dev806+ga2bcb5657",
    "engram_config": {"disk_offload": true},
    "ple_table_approx_gb": 54,
    "runtime_model_gib": 72.89,
    "kv_cache_token_capacity_reported": 847598,
    "single_stream_approx_tok_s": 22,
    "two_stream_aggregate_tok_s": 37.9,
    "long_input_ttft_s": {"approx_30k": 14.0, "approx_60k": 27.9},
    "one_moe_explanation_example": {"tokens": 156, "elapsed_s": 7.5}
  },
  "fp4_local_attempt": {
    "status": "startup_failed",
    "image": "eugr/spark-vllm-b12x:nightly-20261102",
    "model_id": "nvidia/Qwen3.8-Flash-Next-FP4-Local",
    "error_excerpt": "B12xNvFp4LinearKernel: B12X NVFP4 kernels require a Blackwell 12x device"
  }
}