gpt-oss-120b-Derestricted-mxfp4-mlx

MLX MXFP4 quantization of ArliAI/gpt-oss-120b-Derestricted.

Model Capabilities

• Reasoning: Configurable effort (low/medium/high) via reasoning_effort parameter
• Tool Use: Native function calling support
• Context: 131k tokens

Quantization Details

Matches OpenAI's original MXFP4 quantization scheme:

Component	Bits	Group Size	Format
MLP Experts	4	32	MXFP4
Attention	-	-	Full precision (bfloat16)
Routers	-	-	Full precision (bfloat16)
Embeddings	-	-	Full precision (bfloat16)
LM Head	-	-	Full precision (bfloat16)

Usage

mlx-lm

mlx_lm.chat --model txgsync/gpt-oss-120b-Derestricted-mxfp4-mlx

LM Studio

For full Reasoning Effort support in LM Studio, install via:

lms get txgsync/gpt-oss-120b-derestricted

This downloads the model with the virtual model wrapper that enables the Reasoning Effort selector.

Notes

• Requires mlx-lm with gpt_oss HF format support
• Quantization matches OpenAI's modules_to_not_convert scheme for optimal quality