jang_config.json

{
  "version": 2,
  "weight_format": "mxtq",
  "profile": "JANGTQ",
  "cache_subtype": "kv",
  "source_model": {
    "name": "Hy3-preview",
    "org": "tencent",
    "architecture": "hy_v3"
  },
  "expert_layout": "per_expert",
  "mxtq_seed": 42,
  "mxtq_bits": {
    "routed_expert": 2,
    "attention": 8,
    "shared_expert": 8,
    "dense_ffn": 8,
    "mtp": 8,
    "embed_tokens": 8,
    "lm_head": 8,
    "norms_router_biases": 16
  },
  "quantization": {
    "method": "affine+mxtq",
    "group_size": 64,
    "bits_default": 2
  },
  "runtime": {
    "bundle_has_mtp": true,
    "mtp_layers": 1,
    "mtp_mode": "preserved_disabled",
    "mtp_status": "MTP tensors are preserved in the bundle, but the first JANG runtime path must use normal autoregressive decode until an accept/reject speculative loop is implemented and tested."
  },
  "bundle_has_mtp": true,
  "mtp_layers": 1,
  "capabilities": {
    "reasoning_parser": "qwen3",
    "tool_parser": "hunyuan",
    "think_in_template": false,
    "supports_tools": true,
    "supports_thinking": true,
    "family": "hy_v3",
    "modality": "text",
    "cache_type": "kv"
  }
}