{
  "format": "neural-download-model-family-v1",
  "id": "nemotron-3-5",
  "primary_packet_id": "nemotron-35-lightning-30b-a3b-b70",
  "name": "Nemotron 3.5",
  "display_name": "Nemotron 3.5 Lightning 30B-A3B",
  "publisher": "NVIDIA (tested GGUF by unsloth)",
  "updated_at": "2026-08-23",
  "summary": "NVIDIA's Nemotron 3.5 Lightning, a 30B hybrid Mamba mixture-of-experts with about 3B active per word. A fast general assistant that runs on one Arc Pro B70 with stock software.",
  "architecture": {
    "class": "nemotron_h_moe",
    "model_type": "nemotron_h_moe",
    "layers": 53,
    "hidden_size": 2688,
    "experts": 128,
    "experts_used": 6,
    "native_context_tokens": 1048576,
    "evidence": "repro/nemotron-35-lightning-30b-a3b-b70/README.md"
  },
  "weight_revisions": [
    {
      "id": "nemotron-3.5-lightning-30b-a3b",
      "label": "Nemotron 3.5 Lightning 30B-A3B",
      "role": "measured model",
      "model_manifest": "repro/nemotron-35-lightning-30b-a3b-b70/model-manifest.json"
    }
  ],
  "transfer_scope": {
    "status": "Only one model artifact and quantization are measured; no cross-quantization, cross-runtime, or later-weight transfer is claimed.",
    "transfers": [
      "the pinned model verification and stock llama.cpp/SYCL one-card recipe",
      "the fixed benchmark and objective-canary protocol"
    ],
    "does_not_transfer": [
      "speed or quality to NVIDIA NVFP4 or another quantization",
      "the one-card result to another runtime or card topology",
      "the measured 0–32K trend to the untested remainder of the native 1M context window"
    ],
    "evidence": "repro/nemotron-35-lightning-30b-a3b-b70/README.md"
  },
  "model_signals": {
    "b70_fit": {
      "band": "high",
      "scope": "local-llm-on-b70",
      "basis": "The tested UD-Q4_K_M artifact fits and serves target-only on one 32 GiB B70; the measured two-card layer split is slower for single-stream decode.",
      "reviewed_at": "2026-08-23"
    },
    "quality_evidence": {
      "band": "scoped-deployment-evidence",
      "scope": "reasoning-off, temperature-zero repeat stability plus arithmetic, copy, and JSON canaries for the exact UD-Q4_K_M llama.cpp/SYCL deployment; not a general model-quality score",
      "evidence": [
        "repro/nemotron-35-lightning-30b-a3b-b70/README.md",
        "experiments/qwen38-27b-b70/data/2026-08-22-neural-download-firstwave-baselines.json"
      ]
    },
    "popularity": {
      "downloads": 107161,
      "likes": 80,
      "captured_at": "2026-08-22T02:47:17Z",
      "repo_id": "unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
      "scope": "third-party GGUF repository snapshot; popularity is a discovery signal, not validation",
      "evidence": "model-intake/catalog.json"
    }
  },
  "dimensions": {
    "model_variant": [
      "Nemotron 3.5 Lightning 30B-A3B"
    ],
    "weight_quantization": [
      "UD-Q4_K_M"
    ],
    "runtime": [
      "llama.cpp SYCL"
    ],
    "cards": [
      1,
      2
    ],
    "split_mode": [
      "single device",
      "layer"
    ],
    "mtp": [
      0
    ],
    "active_context_tokens": [
      0,
      2048,
      4096,
      8192,
      16384,
      24576,
      32768
    ],
    "configured_max_context_tokens": [
      8192
    ],
    "native_context_tokens": [
      1048576
    ],
    "graph": [
      "off"
    ],
    "kv": [
      "f16"
    ]
  },
  "packets": [
    {
      "id": "nemotron-35-lightning-30b-a3b-b70",
      "label": "Nemotron 3.5 Lightning 30B-A3B · UD-Q4_K_M · one B70",
      "revision": "nemotron-3.5-lightning-30b-a3b",
      "quantization": "UD-Q4_K_M",
      "runtime": "llama.cpp SYCL",
      "cards": 1,
      "status": "candidate",
      "evidence_level": "B70-verified",
      "coverage": [
        "decode",
        "prefill",
        "context through 32K",
        "quality",
        "recipe"
      ],
      "projection": {
        "model": "nemotron3.5_lightning_30b_a3b",
        "quant": "Q4_K_M",
        "runtime": "llama_cpp",
        "spec": "none"
      , "prompt_tokens": 128, "output_tokens": 100},
      "manifest": "packages/nemotron-35-lightning-30b-a3b-b70/package.json"
    }
  ],
  "run_measurements": [
    {
      "id": "nemotron-standard-one-card",
      "state": "lab-measured",
      "revision": "nemotron-3.5-lightning-30b-a3b",
      "variant": "UD-Q4_K_M",
      "runtime": "llama.cpp SYCL 9fee29e",
      "config": {
        "cards": 1,
        "split_mode": "single device",
        "mtp": 0,
        "graph": "off",
        "kv": "f16",
        "configured_max_context_tokens": 8192
      },
      "workload": "two fresh-server runs of the 12-prompt suite, up-to-512-token responses, conventional 99-interval median, target-only, cache zero",
      "metrics": {
        "decode_tok_s": [
          72.169452,
          72.035976
        ]
      },
      "quality": "reasoning-off 8x repeat stability, arithmetic, exact copy, and JSON-schema canaries passed",
      "evidence": "repro/nemotron-35-lightning-30b-a3b-b70/README.md"
    },
    {
      "id": "nemotron-layer-split-two-card",
      "state": "lab-measured",
      "revision": "nemotron-3.5-lightning-30b-a3b",
      "variant": "UD-Q4_K_M",
      "runtime": "llama.cpp SYCL 9fee29e",
      "config": {
        "cards": 2,
        "split_mode": "layer",
        "mtp": 0,
        "graph": "off",
        "kv": "f16",
        "configured_max_context_tokens": 8192
      },
      "workload": "same serving protocol as the one-card operating point, split-mode layer on two B70 cards",
      "metrics": {
        "decode_tok_s": [
          69.445407,
          69.485984
        ]
      },
      "quality": "5/5 canaries passed; measured only as a topology comparison",
      "evidence": "repro/nemotron-35-lightning-30b-a3b-b70/README.md"
    }
  ],
  "series_measurements": [
    {
      "id": "nemotron-context-depth",
      "state": "lab-measured",
      "revision": "nemotron-3.5-lightning-30b-a3b",
      "variant": "UD-Q4_K_M",
      "runtime": "llama.cpp SYCL 9fee29e",
      "config": {
        "cards": 1,
        "split_mode": "single device",
        "mtp": 0,
        "graph": "off",
        "kv": "f16"
      },
      "workload": "llama-bench pp2048 plus tg128 at each depth, flash attention on, five repetitions; raw-engine shape metric",
      "axis": "active_context_tokens",
      "points": [
        {
          "x": 0,
          "decode_tok_s": 73.008586,
          "prefill_tok_s": 1260.029381
        },
        {
          "x": 2048,
          "decode_tok_s": 72.420657,
          "prefill_tok_s": 1166.922303
        },
        {
          "x": 4096,
          "decode_tok_s": 71.846756,
          "prefill_tok_s": 1153.135384
        },
        {
          "x": 8192,
          "decode_tok_s": 70.6952,
          "prefill_tok_s": 1145.645326
        },
        {
          "x": 16384,
          "decode_tok_s": 68.557408,
          "prefill_tok_s": 1106.136083
        },
        {
          "x": 24576,
          "decode_tok_s": 66.552475,
          "prefill_tok_s": 1102.582316
        },
        {
          "x": 32768,
          "decode_tok_s": 64.622975,
          "prefill_tok_s": 1018.926677
        }
      ],
      "evidence": "repro/nemotron-35-lightning-30b-a3b-b70/nemotron-35-lightning.sweep.json"
    }
  ],
  "views": [
    {
      "id": "nemotron-card-count",
      "title": "One card beats layer split",
      "subtitle": "Same UD-Q4_K_M serving protocol; the two-card layer split is measured, not estimated",
      "x_label": "B70 cards",
      "discrete": true,
      "metrics": [
        "decode_tok_s"
      ],
      "series": [
        {
          "label": "serving median",
          "measurement_ids": [
            "nemotron-standard-one-card",
            "nemotron-layer-split-two-card"
          ],
          "x_from": "config.cards"
        }
      ]
    },
    {
      "id": "nemotron-context",
      "title": "Context-depth profile",
      "subtitle": "One B70 · UD-Q4_K_M · f16 KV · raw llama-bench; native context beyond 32K remains unmeasured",
      "x_label": "active context tokens",
      "metrics": [
        "decode_tok_s",
        "prefill_tok_s"
      ],
      "series": [
        {
          "label": "UD-Q4_K_M",
          "measurement_ids": [
            "nemotron-context-depth"
          ]
        }
      ]
    }
  ],
  "coverage_views": [],
  "family_closures": [
    {
      "selectors": {
        "revision": "nemotron-3.5-lightning-30b-a3b",
        "active_context_tokens": ">32768"
      },
      "state": "missing",
      "reason": "The artifact declares a 1,048,576-token native window, but this package has measured context-depth evidence only through 32,768 tokens.",
      "evidence": "repro/nemotron-35-lightning-30b-a3b-b70/README.md"
    }
  ],
  "estimates": []
}