{
  "schema_version": 1,
  "_comment": "Qwen3.5 (qwen3_5) GatedDeltaNet hybrid TEXT decoder for v81. Most layers are a recurrent gated delta-rule linear attention (+ short conv state), 1-in-4 is gated softmax attention (partial-rotary M-RoPE, output gate). Decode-over-prompt seed; fp16 first (strict greedy parity). Vision tower + MTP not exported (text-LLM path).",
  "model": {
    "name": "qwen3.5-2b-1024",
    "family": "llm",
    "dsp_arch": "v81",
    "tokenizer_pre": "qwen2"
  },
  "params": {
    "hidden": 2048,
    "vocab": 248320,
    "n_layers": 24,
    "max_ctx": 1024,
    "kv_dim": 512,
    "head_dim": 256,
    "rope_theta": 10000000.0,
    "eos_token_id": 248044,
    "gated_delta": {
      "q_heads": 8,
      "kv_heads": 2,
      "lin_k_heads": 16,
      "lin_v_heads": 16,
      "lin_k_dim": 128,
      "lin_v_dim": 128,
      "conv_kernel": 4,
      "full_every": 4,
      "rot_dim": 64,
      "attn_gate": true,
      "full_attn_idx": [
        3,
        7,
        11,
        15,
        19,
        23
      ]
    }
  },
  "artifacts": {
    "contexts": {
      "decode": {
        "bin": "qwen352b_decode_f16.bin"
      },
      "lmhead": {
        "bin": "qwen352b_lmhead_f16.bin"
      }
    },
    "embed": "qwen352b_embed_f16.bin",
    "tokenizer": "tokenizer.json"
  },
  "plan": {
    "steps": [
      {
        "host": "qwen3_5_generate",
        "params": {
          "decode": "qwen352b_decode_f16",
          "lmhead": "qwen352b_lmhead_f16",
          "max_new": 64,
          "neg": -50000
        }
      }
    ]
  }
}