Huihui Qwen3.8 27B Coder390 EfficientThink abliterated — oQ8e-mtp (MLX, Apple Silicon)
MLX build of huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated,
quantized from the BF16 source with oMLX oQ8e (imatrix-calibrated 8-bit). The MTP head is preserved for faster decoding.
⚠️ Uncensored model. Refusal behaviour was removed upstream by abliteration (huihui-ai). 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: 27.8B dense Qwen3.8 coder fine-tune (nerkyor Coder390 EfficientThink), abliterated by huihui-ai. 262k context.
- Quantization: oMLX oQ8e — 8-bit affine (group 64), imatrix-calibrated; MTP layers kept.
- Text-only: language model weights only (no vision encoder).
- Size: ~29 GB (source BF16: ~56 GB).
- Runtime: requires oMLX ≥ 0.7.1 (oQ format + Qwen3.5-family MTP). 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 ‖ oQ8e), mean | 0.00115 nat/token |
| KL median / p99 | 0.00052 / 0.0106 |
| Top-1 token agreement | 98.6 % |
| Perplexity BF16 → oQ8e | 6.181 → 6.182 |
Benchmarks (oMLX accuracy suite, thinking off, deterministic)
| Benchmark | Samples | Score |
|---|---|---|
| MMLU-Pro | 300 | 64.7 |
| GSM8K | 100 | 95.0 |
| HumanEval | 164 (full) | 90.8 |
| MBPP | 200 | 80.0 |
| LiveCodeBench | 100 | 68.0 |
| SafetyBench | 300 | 81.7 |
Speed (Mac Studio M3 Ultra, 512 GB, oMLX 0.7.1)
| Prefill, ~29k-token prompt (cold) | ~73 s (~395 tok/s) |
| Same prefix again (prefix cache) | ~1.3 s |
| Decode (with MTP) | ~35 tok/s |
Tool calling (Anthropic /v1/messages / OpenAI /v1/chat/completions, 5 runs each) |
5/5 / 4/5 |
Single-run measurements; numbers vary with context length and settings.
Usage (oMLX)
Place the folder in your oMLX model directory and 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-Qwen3.8-27B-Coder390-EfficientThink-abliterated-oQ8e-mtp", "messages": [{"role": "user", "content": "Write a Python function that parses a CSV line."}]}'
Memory: ~29 GB weights plus KV cache (~96 KB/token on the 24 full-attention layers → ~25 GB at 262k).
Credits & license
- Base fine-tune: nerkyor —
nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2. - Abliteration: huihui-ai —
huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated. - Quantization: oMLX oQ.
- License: Apache-2.0, inherited from upstream.