lemuralabs/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-abliterated-6bit-mlx

<p align="center"> <img src="logo.png" alt="Lemura Labs" width="110"/> </p>

<p align="center">

</p>

Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-abliterated-6bit-mlx

## THIS IS A TEXT-ONLY MODEL — NO VISION The upstream abliteration pass stripped the vision tower. For vision-capable Qwen 3.6 27B Opus-Distill MLX quants, see our parallel repos at huggingface.co/lemuralabs (look for repos without -abliterated in the name).

6-bit affine MLX quantization of an abliterated Qwen 3.6 27B Claude-Opus reasoning distill, by the Lemura Labs team.

Standard 6-bit MLX quantization. Effectively lossless on most reasoning benchmarks vs BF16. The recommended pick when you have the RAM headroom.

TL;DR

Lineage

Qwen/Qwen3.6-27B (Qwen Team — base pretrain)
 │
 ▼
TeichAI/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2 (TeichAI — Claude-Opus reasoning distill)
 │
 ▼
abliterated (refusal-ablated) via OBLITERATUS v0.1.2 (multi-direction SVD, BF16)
 │
 ▼
this repo — 6-bit affine, MLX format (Lemura Labs team — quantization)

Direct upstream links:

• —Foundation: Qwen/Qwen3.6-27B
• —Reasoning distill: TeichAI/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2
• —Abliteration tool: OBLITERATUS (multi-direction SVD, 6 directions, 3 refinement passes, λ=0.08)
• —Quantization tool: mlx-lm + mlx-optiq for OptiQ variants

Use it

mlx-lm (recommended)

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("lemuralabs/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-abliterated-6bit-mlx")
prompt = "Explain the difference between SSM and softmax attention in three sentences."
out = generate(model, tokenizer, prompt=prompt, max_tokens=400)
print(out)

Chat template

messages = [
 {"role": "system", "content": "You are a helpful, candid reasoning assistant."},
 {"role": "user", "content": "Plan a 3-day Tokyo itinerary for a foodie."},
]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
print(generate(model, tokenizer, prompt=prompt, max_tokens=600))

CLI

mlx_lm.generate --model lemuralabs/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-abliterated-6bit-mlx --prompt "Hello" --max-tokens 256

Quantization details

• —Source weights: BF16 abliterated checkpoint (28 shards, ~57 GB) derived from TeichAI/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2 via OBLITERATUS multi-direction SVD ablation (preserves coherence; KL drift = 0.149 from base).
• —Quantization scheme: Affine 6-bit, group size 64 (mlx-lm default).
• —Group size: 64.
• —Calibration corpus: mlx-lm calibration_v5 (~427 KB English text, used for OptiQ sensitivity ranking; uniform/affine variants do not require calibration).
• —Sanity check: forward perplexity on held-out calibration text within 1–3% of next-higher-precision sibling.

Architecture notes

The Qwen 3.6 27B family uses a hybrid attention stack — 4 GatedDeltaNet (linear-attention/SSM) layers followed by 1 full-softmax-attention layer, repeated 16× for 64 total layers, 5120 hidden, 248K vocab, 262K context. The SSM kernels lack a VJP path in MLX, so backward-pass-based quant methods (DWQ, dynamic quant) cannot be applied here — OptiQ's forward-only sensitivity approach is the only calibration-aware option that works on this architecture. That's why the OptiQ variants exist.

Behavior caveats

• —Text-only — no vision. The abliteration pipeline (OBLITERATUS) ran on the LM tower and stripped the ViT. For vision-capable quants of the same Opus-Distill v2 lineage, use our parallel non-abliterated repos at huggingface.co/lemuralabs (any repo without -abliterated in the name).
• —This is an abliterated model — refusal directions were surgically removed from the parent. It will answer prompts the parent would refuse. Use responsibly and within applicable law.
• —Quantization preserves abliteration: the refusal rate measured at BF16 (~35% from a 100% baseline) stays in that range across our quants.

Credits

License

Apache-2.0, inherited from the foundation and distill upstream.

_{Need a hosted endpoint, custom quant, or larger-scale inference? — multi-provider LLM routing for the Indian developer ecosystem.}

3.3–3.7× faster decoding with DFlash (lossless, MLX)

This MLX build supports lossless block-diffusion speculative decoding via [DFlash](https://huggingface.co/z-lab/Qwen3.6-27B-DFlash) in mlx_vlm — no requantization, no model changes. On an Apple M4 Max we measured 3.38× (8-bit) and 3.67× (bf16) decode speedups with byte-identical output; other MLX quants of this model should see a similar ~3×.

python3 -m mlx_vlm generate \
 --model lemuralabs/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-abliterated-6bit-mlx \
 --draft-model z-lab/Qwen3.6-27B-DFlash --draft-kind dflash \
 --prompt "Write a merge function for two sorted lists in Python." --max-tokens 256

• —Requires mlx_vlm ≥ 0.5.0 and access to the gated drafter `z-lab/Qwen3.6-27B-DFlash` (one-click "Agree and access").
• —Accelerates the text path only (vision is unaffected); adds ~3.9 GB for the drafter.
• —Acceptance ≈ 8.95 tokens/round (block size 16); the target runs ~10× fewer forward passes.
• —Full write-up & benchmarks: [] · see also `DFLASH_SPECULATIVE_DECODING.md`.