gpt-oss-120b

66ton99/gpt-oss-120b

This repository contains a local MLX repack of openai/gpt-oss-120b derived from the LM Studio GGUF files.

What this is

• Source: lmstudio-community/gpt-oss-120b-GGUF
• Output: MLX safetensors shard set compatible with mlx-lm
• Conversion style: format repack, not full dequantize/requantize
• Final path in LM Studio: 66ton99/gpt-oss-120b

Weight format

The source GGUF is mixed-format:

• Attention, embeddings, and lm head are stored as 8-bit quantized tensors
• MoE expert weights are stored as MXFP4 tensors
• Norms and routing/bias tensors remain in floating point

The MLX config mirrors that layout:

• group_size: 32, bits: 8, mode: affine for Q8 tensors
• per-layer group_size: 32, bits: 4, mode: mxfp4 for MoE expert tensors

Model details

• Architecture: GPT-OSS MoE
• Layers: 36
• Hidden size: 2880
• Attention heads: 64
• KV heads: 8
• Local experts: 128
• Experts per token: 4
• Context length: 131072

Files

• config.json - MLX-compatible model config with mixed quantization metadata
• model-00001-of-00013.safetensors ... model-00013-of-00013.safetensors - model weights
• model.safetensors.index.json - shard map
• tokenizer.json, tokenizer_config.json, special_tokens_map.json, chat_template.jinja - tokenizer assets

Usage

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("66ton99/gpt-oss-120b")
prompt = "Explain why the sky is blue."
response = generate(model, tokenizer, prompt=prompt, verbose=True)
print(response)

Notes

• This is a conversion artifact, not a retrained checkpoint.
• The weights were repacked from the LM Studio GGUF source into MLX safetensors.