{
  "retrieved_at": "2026-09-23",
  "scope": "Primary/public sources for Ornith-1.5 MoE pruning, quantization, language-routing caveats, and artifact identity. No weights downloaded or executed.",
  "sources": [
    {
      "id": "ornith-35b-card",
      "kind": "official model card",
      "url": "https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B",
      "supports": ["official 35B-A3B base", "MoE family context", "official lineage"],
      "observed": "Public card for the unpruned base; config fetched separately showed Qwen3_5MoeForConditionalGeneration, 256 experts, top-8, 40 layers, MTP 1."
    },
    {
      "id": "ornith-9b-card",
      "kind": "official model card",
      "url": "https://huggingface.co/ornith-ai/Ornith-1.5-9B",
      "supports": ["dense 9B baseline", "official serving/benchmark conditions"],
      "observed": "Card calls this the lightweight dense model and estimates about 19 GB BF16. This is a baseline, not the 35B MoE."
    },
    {
      "id": "reap-repo",
      "kind": "official method repository",
      "url": "https://github.com/CerebrasResearch/reap",
      "supports": ["REAP saliency", "structural expert pruning", "official CLI entrypoints", "router renormalization update"],
      "observed": "README documents pruning-cli.sh, calibration/evaluation flags, and the router-weighted activation criterion."
    },
    {
      "id": "reap-paper",
      "kind": "author paper/arXiv",
      "url": "https://arxiv.org/abs/2510.13999",
      "supports": ["REAP method motivation and evaluation"],
      "observed": "Primary paper for Router-weighted Expert Activation Pruning; results are not Ornith-specific."
    },
    {
      "id": "vllm-reap-example",
      "kind": "official vLLM project example",
      "url": "https://github.com/vllm-project/llm-compressor/tree/main/examples/reap_expert_pruning",
      "supports": ["end-to-end recipe", "Qwen/Moonlight baselines", "25/50% reference results"],
      "observed": "Example uses Qwen3-30B-A3B and Moonlight-16B-A3B; it is a method reference, not Ornith evidence."
    },
    {
      "id": "ttimms-repo",
      "kind": "artifact author repository",
      "url": "https://github.com/t-timms/ornith-nvfp4",
      "supports": ["Ttimms identity", "REAP-50 pipeline", "MTP/vision surgery", "NVFP4 serving/evaluation"],
      "observed": "README identifies t-timms and documents 256→128, MTP strip, vision strip for text-only NVFP4, and vLLM/SM120 evaluation."
    },
    {
      "id": "ttimms-bf16",
      "kind": "artifact author model card",
      "url": "https://huggingface.co/Ttimms/Ornith-1.5-35B-A3B-REAP-50-bf16",
      "revision": "fc562b4c13562893c58a7a07e4ff00f46fe6ba1c",
      "supports": ["published REAP-50 BF16 source", "128 experts", "MTP disabled", "source for custom quants"],
      "observed": "Card says 256→128, router renormalization, MTP 0, vision retained in BF16 source."
    },
    {
      "id": "ttimms-nvfp4",
      "kind": "artifact author model card",
      "url": "https://huggingface.co/Ttimms/Ornith-1.5-35B-A3B-REAP-50-NVFP4A16",
      "revision": "a38f29db4d12b4cfffcdcfec5dfad7244cf7097c",
      "supports": ["12.47 GiB NVFP4 artifact", "SM120/vLLM target", "author-reported coding and SWE-bench results"],
      "observed": "Quantized derivative of the REAP-50 source; do not transfer its numbers to BF16, GGUF, or 9B."
    },
    {
      "id": "oxfrug-20pct",
      "kind": "artifact author model card",
      "url": "https://huggingface.co/oxfrug/Ornith-1.5-35B-A3B-REAP-20pct",
      "revision": "1407285ab59071d81cb64a6f4d9fac8d6a670a75",
      "supports": ["exact 20% candidate identity", "205 experts", "top-8 preserved", "MTP/vision/shared expert handling", "recipe and limits"],
      "observed": "Card reports 256→205 (51 removed), router renormalization, MTP/vision/embeddings/lm_head untouched, and explicitly warns that published proxy scores are not runs of this checkpoint."
    },
    {
      "id": "sleepyeldrazi-20b",
      "kind": "artifact author model card",
      "url": "https://huggingface.co/sleepyeldrazi/Ornith-1.5-REAP-20B-A3B-GGUF",
      "revision": "bee3dcd493946c5b613493c9db7908b8017ede3c",
      "supports": ["20B total artifact distinction", "132 experts", "recovery LoRA", "domain-specific PPL and limits"],
      "observed": "Card says 35.9B→20.3B, 256→132, 48.4% sparsity, recovery fine-tuning, and general/multilingual behavior untested."
    },
    {
      "id": "mchen04-reap192",
      "kind": "artifact author model card",
      "url": "https://huggingface.co/mchen04/Ornith-1.5-35B-A3B-REAP192-mxfp4-MLX",
      "revision": "ad9d71002ca26b08d769492f733892087c057242",
      "supports": ["25% retention variant", "top-k 8→7", "MLX/mxfp4 constraints"],
      "observed": "Card documents 256→192, top-k 7, no fine-tuning, and Apple Silicon measurements."
    },
    {
      "id": "optiq-reap19b",
      "kind": "community/MLX artifact card",
      "url": "https://huggingface.co/mlx-community/Ornith-1.5-35B-A3B-OptiQ-4bit-REAP-19B",
      "revision": "d2930458b7e8ff4799d5e17ea1ef23818746361e",
      "supports": ["independent 50% OptiQ derivative", "warning that Ornith was not benchmarked separately"],
      "observed": "Card says 128/256 experts and explicitly transfers quality evidence from a Qwen3.6 sibling."
    },
    {
      "id": "multilingual-routing",
      "kind": "author paper/arXiv",
      "url": "https://arxiv.org/abs/2510.04694",
      "supports": ["language-routing evidence from other MoEs", "shared-vs-language-correlated routing caveat"],
      "observed": "Paper reports language-specific routing in early/late layers and cross-lingual alignment in middle layers; no Ornith-specific result."
    },
    {
      "id": "vocabulary-trimming",
      "kind": "peer-reviewed primary paper",
      "url": "https://aclanthology.org/2023.findings-emnlp.981/",
      "supports": ["vocabulary compression distinction from expert pruning"],
      "observed": "Vocabulary trimming results cover other multilingual LMs and seven languages; not a direct Ornith recipe."
    },
    {
      "id": "tofu",
      "kind": "author paper/arXiv",
      "url": "https://arxiv.org/abs/2401.06121",
      "supports": ["unlearning evaluation caveat"],
      "observed": "TOFU argues that refusal/behavior change is not sufficient evidence of behaving as if the data had never been learned."
    }
  ]
}
