hakunokina/Huihui-Qwen3-Coder-Next-abliterated-oQ4e

Huihui Qwen3-Coder-Next abliterated — oQ4e (MLX, Apple Silicon)

MLX build of huihui-ai/Huihui-Qwen3-Coder-Next-abliterated (an abliterated edition of Qwen/Qwen3-Coder-Next), quantized from the BF16 source with oMLX oQ4e (imatrix-calibrated 4-bit). Until now this model had only plain MLX 4/8-bit builds and no oQ build.

⚠️ Abliterated model. Refusal behaviour 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-Coder-Next, ~80B total / ~3B active MoE (512 experts, top-10, plus a shared expert), 48 layers, hybrid Gated DeltaNet + gated attention (full attention every 4th layer), 262k context. A non-thinking coding/agent model.
• Quantization (oQ4e): affine 4-bit, group size 64, with 314 sensitive tensors at 8-bit, 36 at 5-bit; imatrix-calibrated on oMLX's oqe_code_multilingual set. Effective 4.61 bits/weight.
• No MTP head (the source has none).
• Size: 46 GB (43 GiB) (source BF16: ~159 GB). Other size: oQ6e (65 GB).
• Runtime: built and tested with oMLX 0.7.3. Not tested with plain mlx-lm / LM Studio.

Quality vs. BF16 source

Full-vocabulary KL against the BF16 source, same token streams; 64 sequences × 512 tokens (32.7k scored tokens): code, English prose, Turkish.

Metric	Value
KL(BF16 ‖ oQ4e), mean	0.119 nat/token
KL median / p99	0.015 / 1.45
Top-1 token agreement	88.8 %
Perplexity BF16 → oQ4e	6.128 → 6.365 (+3.9 %)

Qwen3-Next's hybrid layout (Gated DeltaNet + gated attention, 512 tiny experts) is quantization-sensitive: at 4 bits perplexity rises +3.9 % and top-1 agreement drops to 88.8 %. Tool calling still passed every run, but for long or subtle coding work prefer the oQ6e build; choose oQ4e when memory is the constraint (64 GB Macs).

No benchmark suite was run on this build; quality is reported as divergence from the source.

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

Prefill, ~27k-token prompt (model loaded, cold cache)	14 s (1,880 tok/s)
Same prefix again (prefix cache)	~0.70 s
Decode	~72 tok/s
Tool calling (Anthropic /v1/messages / OpenAI /v1/chat/completions, 5 runs each)	5/5 / 5/5

The very first request (~27k tokens, including model load) took ~16 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": "Huihui-Qwen3-Coder-Next-abliterated-oQ4e", "messages": [{"role": "user", "content": "Write a Python function that parses a CSV line."}]}'

Suggested sampling (Qwen3-Coder recommendation): temperature 0.7, top_p 0.8, top_k 20, repetition_penalty 1.05.

The source's tool-call helpers (qwen3coder_tool_parser_vllm.py, qwen3_coder_detector_sgl.py) are included for vLLM/SGLang users; oMLX does not need them.

Credits & license

• Base model: Qwen team — Qwen/Qwen3-Coder-Next.
• Abliteration: huihui-ai — huihui-ai/Huihui-Qwen3-Coder-Next-abliterated.
• Quantization: oMLX oQ.
• License: Apache-2.0, inherited from upstream.