---
language: en
library_name: mlx
pipeline_tag: text-generation
tags:
- mlx
- axquant
- gpt-oss
- moe
license: apache-2.0
base_model: openai/gpt-oss-120b
---

# AX-gpt-oss-120b-MLX-AXQ-6bit

AXQuant affine re-pack of mlx-community/gpt-oss-120b-MXFP4-Q4 for Apple Silicon MLX.

| Property | Value |
| --- | --- |
| Product | AXQ 6-bit (agent-coding manual recipe, no 4-bit trunk) |
| Measured total BPW | 6.577 |
| Architecture | GptOssForCausalLM (MoE, no MTP) |
| Source | mlx-community/gpt-oss-120b-MXFP4-Q4@bce781bef0f2fc85ed4e575af74054f5aad73ddd |
| Upstream | openai/gpt-oss-120b |
| Plan | plan-manual agent-coding: experts 6-bit, attention 8-bit, no 4-bit |
| Runtime | MLX-LM |

## Checkpoint Tier 1

Certified on host df-macbookpro-m5 with AXQuant 1.6.1 development suites
(agent-coding + general, seed 20260728, max tokens 64) vs the matched MXFP4-Q4
reference. MTP Tier 2 is not applicable (no MTP).

## Load

\`\`\`bash
pip install mlx-lm
python -m mlx_lm.generate --model AutomatosX/AX-gpt-oss-120b-MLX-AXQ-6bit --prompt Hello
\`\`\`

## Notes

- Converted with AXQUANT_FORCE_CPU=1 after Metal GPU timeouts on large re-pack.
- Size ratio vs MXFP4-Q4 is ~1.54 (within the 6-bit max 1.55 gate).
- Higher-fidelity 6-bit product layout (storage-adjusted BPW ~6.58), not uniform 6.0.

## Modalities (capability-gated)

Text checkpoint Tier 1 does **not** imply vision or audio quality. `Vision present=true` on a pack is not a quality pass.

| Modality | Claim | Supported | Reason |
| --- | --- | --- | --- |
| Vision | `not-applicable` | `false` | vision not supported (no tower config and no sidecar weights) |
| Audio | `not-applicable` | `false` | audio not supported (no tower config and no sidecar weights) |