Interface: ModelConfig

Properties

expert _ dtype

• Optional expert_dtype: string

Some MoEs store their experts at a narrower precision than the rest of the model and declare it here, outside quantization_config (e.g. DeepSeek-V4 is quant_method: "fp8" for attention but expert_dtype: "fp4" for the experts, which dominate the parameter count).

Defined in

packages/hub/src/lib/parse-safetensors-metadata.ts:586

---

quantization _ config

• Optional quantization_config: QuantizationConfig

Defined in

packages/hub/src/lib/parse-safetensors-metadata.ts:579

---

text _ config

• Optional text_config: { quantization_config?: QuantizationConfig } & Pick<ModelConfig, "expert_dtype">

Defined in

packages/hub/src/lib/parse-safetensors-metadata.ts:580

Update on GitHub