Qwen3.5-2B OneCompression 4-bit MLX
Text-only MLX checkpoint produced from Qwen/Qwen3.5-2B for AyaneSDK's
on-device conversation example.
- OneCompression GPTQ with quantization-error propagation (QEP)
- 4-bit weights, group size 128
- 256 Japanese dialogue calibration samples of 512 tokens
- 186 quantized linear layers
- Token embedding quantized separately to asymmetric MLX 4-bit
- Vision weights are not included
The packed GPTQ-v1 linear weights were converted losslessly to MLX's
row-major affine representation. The model configuration retains an
onecompression_source_quantization audit record.
Source
Base model: Qwen/Qwen3.5-2B
Quantizer: FujitsuResearch/OneCompression