Qwen3-235B-A22B abliterated — oQ4e (MLX, Apple Silicon)

MLX build of null-space/Qwen3-235B-A22B-abliterated-FP8 (an abliterated edition of Qwen/Qwen3-235B-A22B), packed with oMLX oQ4e (imatrix-calibrated 4-bit) for Apple Silicon. Until now this model had no MLX/oQ build.

⚠️ Abliterated model. The refusal direction was removed by the upstream author at the weight level. It will follow instructions it would otherwise refuse. You are responsible for how you use it and for complying with applicable law.

What this build is

• Model: Qwen3-235B-A22B, ~235B total / ~22B active MoE (128 experts, top-8), 94 layers. Hybrid thinking model (/think and /no_think switches). 32k native context (extendable to ~128k with YaRN).
• Quantization (oQ4e): affine 4-bit, group size 64, with 378 sensitivity-selected tensors at 8-bit; imatrix-calibrated on oMLX's oqe_code_multilingual set (all 128 experts received calibration data). Effective 4.64 bits/weight.
• Source: the upstream FP8 (compressed-tensors) checkpoint, ~236 GB.
• No MTP head (the source has none).
• Size: ~136 GB (127 GiB).
• Runtime: built and tested with oMLX 0.7.3. Not tested with plain mlx-lm / LM Studio.

Quality

• KL divergence against the source was not measured: our KL harness cannot read the FP8 compressed-tensors format, and the BF16 reference (null-space/Qwen3-235B-A22B-abliterated, ~438 GB) was not downloaded for this upload.
• No accuracy benchmarks were run. See the upstream card for the abliteration evaluation.
• Sanity checks: coherent chat and tool calling (below).

Speed (Mac Studio M3 Ultra, 512 GB, oMLX 0.7.3)

The very first request (~27k tokens, including model load) took ~123 s. Single-run measurements; numbers vary with context length and settings.

Usage (oMLX)

Place the folder in your oMLX model directory and reload models (or restart oMLX). It is served on both the OpenAI and Anthropic APIs:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-235B-A22B-abliterated-oQ4e", "messages": [{"role": "user", "content": "Write a short story opening in two sentences."}]}'

Suggested sampling (Qwen3 recommendation): thinking temperature 0.6, top_p 0.95, top_k 20; non-thinking temperature 0.7, top_p 0.8, top_k 20. A presence_penalty around 1.0 helps against repetition in quantized builds.

Memory: ~136 GB of weights. Fits a 192 GB Mac at moderate contexts; 256 GB+ recommended for long contexts.

Credits & license

• Base model: Qwen team — Qwen/Qwen3-235B-A22B.
• Abliteration and FP8: null-space — null-space/Qwen3-235B-A22B-abliterated-FP8.
• Quantization: oMLX oQ.
• License: Apache-2.0, inherited from upstream.