{
  "schema": "aiinfra.dgx-spark-model-evidence.v1",
  "generatedAt": "2026-08-15",
  "hardware": "NVIDIA DGX Spark / GB10, single node unless explicitly stated",
  "scope": "Official documented workflows and reproducible community measurements; not an exhaustive compatibility list",
  "disclaimer": "Official workflow does not imply a performance guarantee. Community benchmark does not imply NVIDIA endorsement or the same result on another software stack.",
  "records": [
    {
      "id": "official-qwen36-ollama-nvfp4",
      "model": "Qwen3.6 35B-A3B",
      "workload": "Local coding agent inference",
      "evidenceStatus": "official_workflow",
      "engine": "Ollama",
      "format": "NVFP4",
      "context": "256K model context listed",
      "result": "Official playbook lists an approximately 22GB model package",
      "caveat": "Package size is not total runtime memory use",
      "sourceCheckedAt": "2026-08-15",
      "sourceUrl": "https://build.nvidia.com/spark/cli-coding-agent"
    },
    {
      "id": "official-qwen36-ollama-bf16",
      "model": "Qwen3.6 35B-A3B",
      "workload": "Local coding agent inference",
      "evidenceStatus": "official_workflow",
      "engine": "Ollama",
      "format": "BF16",
      "context": "256K model context listed",
      "result": "Official playbook lists an approximately 71GB model package",
      "caveat": "Unified memory must also hold runtime, KV cache, system and display use",
      "sourceCheckedAt": "2026-08-15",
      "sourceUrl": "https://build.nvidia.com/spark/cli-coding-agent"
    },
    {
      "id": "official-qwen36-llamacpp-q4",
      "model": "Qwen3.6 35B-A3B MTP",
      "workload": "OpenAI-compatible local inference API",
      "evidenceStatus": "official_workflow",
      "engine": "llama.cpp",
      "format": "GGUF Q4_K_XL",
      "context": "Server auto-fit; official notes recommend at least 32K and preferably 100K+ for agentic coding",
      "result": "Official build, serve, health-check and MTP recipe",
      "caveat": "Context and concurrent slots materially change KV-cache use",
      "sourceCheckedAt": "2026-08-15",
      "sourceUrl": "https://build.nvidia.com/spark/llama-cpp/instructions"
    },
    {
      "id": "official-nemotron3",
      "model": "Nemotron-3 Nano / Nemotron-3 Super",
      "workload": "OpenAI-compatible inference service",
      "evidenceStatus": "official_workflow",
      "engine": "vLLM or TensorRT-LLM",
      "format": "Model-specific; Super includes an NVFP4 path",
      "context": "Model and recipe dependent",
      "result": "Official single-Spark deployment recipes",
      "caveat": "Container tags, plugins and model access must match the recipe",
      "sourceCheckedAt": "2026-08-15",
      "sourceUrl": "https://build.nvidia.com/spark/nemotron/overview"
    },
    {
      "id": "official-llama31-70b-qlora",
      "model": "Llama 3.1 70B",
      "workload": "Parameter-efficient fine-tuning",
      "evidenceStatus": "official_workflow",
      "engine": "PyTorch",
      "format": "LoRA / QLoRA recipe",
      "context": "Recipe example uses sequence length 2048",
      "result": "Official 70B LoRA and QLoRA scripts",
      "caveat": "This is not full-parameter 70B training and not training a foundation model from scratch",
      "sourceCheckedAt": "2026-08-15",
      "sourceUrl": "https://build.nvidia.com/spark/pytorch-fine-tune/instructions"
    },
    {
      "id": "official-qwen3-8b-sft",
      "model": "Qwen3-8B",
      "workload": "Supervised fine-tuning",
      "evidenceStatus": "official_workflow",
      "engine": "NeMo AutoModel",
      "format": "Full SFT example",
      "context": "Official smoke example uses packed sequence size 1024",
      "result": "Official Spark-specific short-run validation recipe",
      "caveat": "A 20-step smoke run validates the pipeline, not final model quality",
      "sourceCheckedAt": "2026-08-15",
      "sourceUrl": "https://build.nvidia.com/spark/nemo-fine-tune/instructions"
    },
    {
      "id": "official-flux-sdxl",
      "model": "FLUX.1 / SDXL",
      "workload": "Text-to-image inference",
      "evidenceStatus": "official_workflow",
      "engine": "TensorRT",
      "format": "FP16 / FP8 / FP4 paths",
      "context": "Not applicable",
      "result": "Official multi-modal inference workflow",
      "caveat": "Requires model access; quantization can affect image quality",
      "sourceCheckedAt": "2026-08-15",
      "sourceUrl": "https://build.nvidia.com/spark/multi-modal-inference"
    },
    {
      "id": "community-qwen36-nvfp4-mtp",
      "model": "Qwen3.6 35B-A3B NVFP4 MTP",
      "workload": "Agent-style LLM inference",
      "evidenceStatus": "community_reproducible_benchmark",
      "engine": "vLLM",
      "format": "NVFP4 with FP8 KV, MTP n3",
      "context": "256K configured; PBM fixed fills at 4K, 50K and 100K",
      "result": "86.3 tok/s at 4K, 78.8 at 50K, 31.5 at 100K",
      "caveat": "Single community box and recipe; a separate 50K agent run reports tool_ok=false",
      "sourceCheckedAt": "2026-08-15",
      "sourceUrl": "https://sparkbench.dev/models/nvidia_qwen3.6-35b-a3b/"
    },
    {
      "id": "community-qwen3-30b-a3b",
      "model": "Qwen3-30B-A3B NVFP4",
      "workload": "Agent-style LLM inference",
      "evidenceStatus": "community_reproducible_benchmark",
      "engine": "vLLM",
      "format": "NVFP4 with FP8 KV",
      "context": "40K configured; PBM fixed fill at 4K",
      "result": "74.2 tok/s at 4K",
      "caveat": "Single community box and recipe; a separate approximately 18K-fill agent run reports tool_ok=false",
      "sourceCheckedAt": "2026-08-15",
      "sourceUrl": "https://sparkbench.dev/models/nvidia_qwen3-30b-a3b/"
    },
    {
      "id": "community-qwen3-coder-next-80b",
      "model": "Qwen3-Coder-Next 80B NVFP4",
      "workload": "Coding-agent inference",
      "evidenceStatus": "community_reproducible_benchmark",
      "engine": "vLLM",
      "format": "NVFP4 with FP8 KV",
      "context": "256K configured; PBM fixed fill at 4K",
      "result": "58.9 tok/s at 4K",
      "caveat": "Single community box and recipe; a separate 50K-fill agent run reports tool_ok=false",
      "sourceCheckedAt": "2026-08-15",
      "sourceUrl": "https://sparkbench.dev/models/saricles_qwen3-coder-next/"
    }
  ]
}