jang_config.json

{
  "version": 2,
  "weight_format": "mxtq",
  "profile": "JANGTQ2",
  "source_model": {
    "name": "Nemotron-3-Nano-Omni-30B-A3B-Reasoning",
    "org": "nvidia",
    "architecture": "nemotron_h",
    "wrapper_arch": "NemotronH_Nano_Omni_Reasoning_V3",
    "modality": "omni"
  },
  "mxtq_seed": 42,
  "mxtq_bits": {
    "attention": 8,
    "shared_expert": 8,
    "mamba_proj": 8,
    "routed_expert": 2,
    "embed_tokens": 8,
    "lm_head": 8
  },
  "quantization": {
    "method": "affine+mxtq",
    "group_size": 64,
    "bits_default": 2
  },
  "hybrid_pattern": "MEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEMEM*EMEMEMEME",
  "capabilities": {
    "reasoning_parser": "deepseek_r1",
    "tool_parser": "nemotron",
    "think_in_template": true,
    "supports_tools": true,
    "supports_thinking": true,
    "family": "nemotron_h",
    "modality": "omni",
    "cache_type": "hybrid"
  },
  "modality": "omni",
  "multimodal_components": [
    "vision_model",
    "sound_encoder",
    "mlp1",
    "sound_projection"
  ]
}