Built for vMLX — the MLX inferencer with KV-cache quantization, prefix-cache reuse, agentic tool calling, and hybrid Conv1d+attention cache support.
Free for macOS · vmlx.net
LFM 2.5 8B-A1B — MXFP8 CRACK
CRACK abliterated · 8-bit microscaling · Hybrid Conv1d + Attention + 32-expert MoE · Reasoning · 8.2 GB
What Is This?
This is LFM 2.5 8B-A1B by Liquid AI — a hybrid
Conv1d + full-attention + 32-expert MoE (24 layers, 8B total / ~1B active, 128k context,
reasoning with <think> traces) that has been:
- CRACK abliterated — refusal behavior removed at the weight level. The model now complies across all task categories instead of refusing, while keeping its knowledge, reasoning, and tool use intact.
- 8-bit microscaling quantized for MLX on Apple Silicon — 8.2 GB.
Results
Evaluated through the Osaurus runtime. Compliance graded via HarmBench-text refusal classifier; MMLU via logit-mode argmax over A/B/C/D token logits (matched on both base and CRACK with identical chat-template rendering).
HarmBench compliance (70 prompts · 10 per category)
| Category | CRACK ASR |
|---|---|
| Chemical / biological | 10 / 10 (100%) |
| Copyright | 10 / 10 (100%) |
| Cybercrime / intrusion | 10 / 10 (100%) |
| Harassment / bullying | 10 / 10 (100%) |
| Illegal | 10 / 10 (100%) |
| Misinformation / disinformation | 10 / 10 (100%) |
| General harmful | 10 / 10 (100%) |
| Overall | 70 / 70 (100%) |
MMLU-228 (57-subject, 4 q per subject)
| Subject area | base | CRACK | Δ |
|---|---|---|---|
| Overall | 52.6% | 52.6% | +0.0pp |
| STEM | 43.1% | 40.3% | -2.8pp |
| Humanities | 46.2% | 51.9% | +5.7pp |
| Social Sciences | 70.8% | 66.7% | -4.1pp |
| Other (medicine, business, …) | 55.4% | 57.1% | +1.7pp |
Features
- 128k context with hybrid cache
- Reasoning — emits
<think>...</think>traces; vMLX's reasoning parser surfaces them inmessage.reasoning_contentand the final answer inmessage.content - Tool calling — Liquid Python-call syntax inside
<|tool_call_start|>...<|tool_call_end|>, parsed by vMLX'slfm2tool parser - Hybrid Conv1d + full-attention + 32-expert MoE (24 layers, ~1B active params)
Usage
Run with vMLX (recommended — supports hybrid cache + reasoning + tools) or
an MLX runtime with lfm2_moe support.
Liquid AI recommends temperature=0.3, min_p=0.15, repetition_penalty=1.05 for general use.
# OpenAI-compatible chat completion
# POST /v1/chat/completions
{
"model": "dealignai/LFM2.5-8B-A1B-MXFP8-CRACK",
"messages": [{"role": "user", "content": "..."}],
"temperature": 0.3, "min_p": 0.15,
"repetition_penalty": 1.05
}
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.