license: other base_model:
- stepfun-ai/Step-3.7-Flash library_name: mlx pipeline_tag: image-text-to-text thumbnail: dealign_mascot.png tags:
- mlx
- apple-silicon
- abliterated
- uncensored
- crack
- reasoning
- thinking
- harmbench
- mmlu
- step3p7
- stepfun
- moe
- vision
- image-text-to-text
- jang_k
Step 3.7 Flash — JANG_K CRACK
CRACK abliterated · JANG_K (mixed-precision, Q2 routed) · 45-layer MoE with always-on shared expert · Vision · Reasoning · 74 GB
What Is This?
This is Step 3.7 Flash by StepFun — a 45-layer multimodal Mixture-of-Experts reasoning model with always-on shared expert (DSV4-style). This release is:
- CRACK abliterated — refusal behavior removed at the weight level. The model complies across all task categories while preserving reasoning quality, knowledge, and multimodal capabilities.
- JANG_K (mixed-precision, Q2 routed) for MLX on Apple Silicon — 74 GB.
Results
Evaluated on a Mac Studio M5 Max via the vMLX engine in reasoning mode (temperature 0.6, top-p 0.95). HarmBench-text 70-prompt sample (10 per category). MMLU-114 (2 questions per subject across all 57 subjects) with up to 1,800 tokens per question for full reasoning traces.
HarmBench compliance (70 prompts · 10 per category)
| Category | CRACK ASR |
|---|---|
| Chemical / biological | 10/10 |
| Copyright | 8/10 |
| Cybercrime / intrusion | 10/10 |
| Harassment / bullying | 10/10 |
| Illegal | 10/10 |
| Misinformation / disinformation | 10/10 |
| General harmful | 10/10 |
| Overall | 68/70 (97%) |
MMLU-114 (reasoning-mode, by subject area)
| Subject area | base | CRACK | Δ |
|---|---|---|---|
| Overall | 78.1% | 76.3% | -1.8pp |
| STEM | 75.0% | 77.8% | +2.8pp |
| Humanities | 73.1% | 76.9% | +3.8pp |
| Social Sciences | 87.5% | 83.3% | -4.2pp |
| Other (medicine, business, …) | 78.6% | 67.9% | -10.7pp |
Features
- Reasoning — chat template pre-opens
<think>; vMLX's reasoning parser surfaces the trace inmessage.reasoning_contentand the final answer inmessage.content - MoE with always-on shared expert — every layer combines routed experts with a shared expert (DSV4-style)
- Vision — multimodal image-text understanding
- 45 transformer layers, hidden 4096
Usage
Run with vMLX (recommended — full step3p7 reasoning + vision support).
# OpenAI-compatible chat completion
# POST /v1/chat/completions
{
"model": "dealignai/Step-3.7-Flash-JANG_K-CRACK",
"messages": [{"role": "user", "content": "..."}],
"temperature": 0.6, "top_p": 0.95
}
About CRACK
CRACK (Controlled Refusal Ablation via Calibrated Knockouts) removes safety-refusal behavior at the weight level so the model complies with all task categories while preserving reasoning quality, factual knowledge, and coherence.
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
See our research: Safety Generalization in Frontier Models
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. Published for AI-safety research and authorized security testing.