[!IMPORTANT] Set your sampler explicitly: temperature=1.0, top_p=0.95, top_k=20. Stop tokens eos_token_id = [248046, 248044]. Reasoning is on by default at xhigh; tiers xhigh (default) / medium / low via the chat template reasoning_effort kwarg; disable with enable_thinking=False. Vision + video and the native MTP draft head are preserved.

Built for vMLX — the MLX inferencer with KV-cache quantization, prefix-cache reuse, agentic tool calling, speculative decoding, and mixed-precision JANG bundles.
_{Free for macOS · vmlx.net}

---

Qwen 3.8 27B — MXFP8 CRACK

CRACK abliterated · JANG 8-bit MXFP8 (MLX) · Dense hybrid GatedDeltaNet + attention · Vision + Video · Reasoning tiers · Tools · Native MTP · ~27 GB

---

What Is This?

This is Qwen/Qwen3.8-27B — a 27B dense hybrid (GatedDeltaNet linear-attention + gated full-attention) vision-language model with video understanding, reasoning-effort tiers, tool calling, and a native Multi-Token-Prediction draft head — that has been:

1. CRACK abliterated — refusal behavior removed at the weight level, so it complies across task categories instead of refusing, while keeping reasoning, vision, tools, and knowledge intact.
2. MXFP8 quantized — 8-bit MXFP8 MLX bundle for Apple Silicon (~27 GB).

Vision, video, reasoning tiers, XML tool-calling, and native MTP speculative decoding are all preserved.

Results

Evaluated through the MLX runtime. HarmBench scored with a strict code/chemistry-aware classifier (only substantive, coherent, on-topic compliance counts). MMLU is the standard 57-subject benchmark in logit mode.

MMLU moves +0.0pp — within run-to-run noise (no subject collapse). Refusal behavior removed; capability, reasoning, vision, tools, and multilingual (EN+ZH) preserved.

Features

• Dense hybrid — GatedDeltaNet linear-attention + gated full-attention, 64 layers.
• Vision + Video — image and video understanding preserved (image-text-to-text).
• Native MTP — the Multi-Token-Prediction draft head is preserved and CRACK-aligned (the draft head is cracked to match the uncensored model, so its drafts track the compliant outputs) for speculative decoding — measured preserved draft acceptance on this quant. Auto-engages at temperature 0 / deterministic sampling.
• Reasoning tiers — on by default at xhigh; xhigh / medium / low via reasoning_effort; <think>…</think>; disable with enable_thinking=False.
• Tool calling — native XML function-call schema preserved.
• Multilingual — English + Chinese.

Usage

from mlx_vlm import load, generate
model, processor = load("dealignai/Qwen3.8-27B-MXFP8-CRACK")
# recommended sampling: temperature=1.0, top_p=0.95, top_k=20; eos [248046, 248044]

Benchmarks — all quant levels

Every quant validated independently: HarmBench harm-240 (thinking-off, strict code/chemistry-aware classifier — only substantive, coherent, on-topic compliance counts) and MMLU (57-subject, logit mode, base vs CRACK on the identical harness).

All four reach 100% HarmBench compliance with MMLU held within a couple of points of base (6D actually improves). Pick by memory budget: 6D best quality, 4D the balance, 2D smallest, MXFP8 reference 8-bit. Base refuses HarmBench by design (not shown — comparison is compliance vs. capability).

About CRACK

CRACK (Controlled Refusal Ablation via Calibrated Knockouts) is dealignai's weight-level method for removing safety-refusal behavior while preserving reasoning quality, coherence, and general capability — so the model complies across task categories instead of refusing. Calibrated per model.

Support dealignai

All models are built from original research and released free.

Support us on Ko-fi — membership gets early access and extras.

Ko-fi · X @dealignai · dealign.ai

---

Disclaimer

This model has had its safety-refusal behavior removed for research purposes. It will follow instructions across all categories without refusing. You are solely responsible for how you use it and for complying with all applicable laws. Base model © Alibaba (Apache-2.0). Published for AI-safety research and authorized security testing.