{
  "model": "razorstrike-v1-GGUF",
  "source_model": "lancejames221b/razorstrike-v1",
  "base_model": "Qwen/Qwen3.6-35B-A3B",
  "build_date": "2026-07-21",
  "format": "GGUF",
  "published_files": [
    {
      "path": "razorstrike-v1-IQ4_XS.gguf",
      "quantization": "IQ4_XS",
      "bpw": 4.37,
      "local_size": "18G"
    },
    {
      "path": "mmproj.gguf",
      "role": "multimodal projector",
      "local_size": "858M"
    }
  ],
  "local_intermediates_not_uploaded": [
    {
      "path": "razorstrike-v1-bf16.gguf",
      "reason": "65G bf16 conversion intermediate"
    },
    {
      "path": "/Volumes/Scratch/mlx-quant-work/holo3-qwopus-agentworld-bf16-hfnaming-rawnorm",
      "reason": "raw-HF-compatible reconstruction used for conversion"
    }
  ],
  "root_cause_fixed": {
    "symptom": "previous GGUF loaded but generated corrupted first-token text",
    "cause": "MLX-derived source checkpoint had +1.0 baked into non-linear-attention text RMSNorm weights; llama.cpp Qwen converter expects raw HF weights and applies +1.0 itself",
    "fixes": [
      "subtract 1.0 from model.language_model.*norm.weight except linear_attn.norm.weight",
      "preserve linear_attn.norm.weight unchanged",
      "preserve model.visual.* norms unchanged",
      "merge switch_mlp gate/up into experts.gate_up_proj",
      "rename switch_mlp.down_proj.weight to experts.down_proj",
      "convert with --no-mtp"
    ]
  },
  "verification": {
    "raw_reconstruction_mlx_smoke": "What is 5 plus 3? -> 8",
    "llama_cpp_text_smoke": "What is 17 times 24? -> 408",
    "llama_cpp_image_smoke": "green square, yellow circle, TEST-42 correctly identified",
    "server_args": "llama-server -m razorstrike-v1-IQ4_XS.gguf --mmproj mmproj.gguf -c 4096 -fit off"
  }
}