[!IMPORTANT] Set your sampler explicitly:
temperature=1.0,top_p=1.0,top_k=20— the poolside vendor defaults documented for Laguna XS 2.1. These values are baked intogeneration_config.jsonandjang_config.json → sampling_defaults, but many inference runtimes ignore those files and fall back to notop_kfilter, which produces incoherent / degraded output on this model. Pass them explicitly at your API request layer or CLI flags.
Laguna-XS 2.1 — JANG_4M CRACK
CRACK abliterated · JANG mixed-precision (MLX affine) · Agentic coding · Reasoning toggle · Tool calling · 262K context · ~18 GB
What Is This?
This is poolside/Laguna-XS-2.1 — a 33B-parameter / ~3B-active Mixture-of-Experts model (40 layers, 256 routed experts top-8 + shared expert, hybrid global + sliding-window attention, 262K context) built for agentic coding — that has been:
- CRACK abliterated — refusal behavior removed at the weight level, so it complies across task categories instead of refusing, while keeping its coding ability, knowledge, and reasoning intact.
- JANG_4M quantized — a balanced mixed-precision affine MLX bundle for Apple Silicon (~18 GB). High-precision attention, shared expert, router, and norms; the routed-expert bulk carries the low-bit weight.
Text-only. Reasoning and tool calling fully preserved.
Results
Evaluated through the vMLX inference engine. HarmBench scored with a strict classifier (rejects loops, empty/template dumps, and thinking-trace leakage). MMLU is the standard 57-subject multiple-choice benchmark in logit mode.
| Metric | Base | CRACK |
|---|---|---|
| MMLU (57-subject) | 73.7% | 73.4% (−0.4) |
| HarmBench-320 (compliance / ASR) | — | 99.4% (318/320) |
Abliteration is essentially capability-neutral — MMLU moves −0.4 points, within run-to-run noise — while removing refusals. The model stays coherent in both direct and reasoning modes.
MMLU by topic (base vs CRACK)
| Category | Base | CRACK | Δ |
|---|---|---|---|
| STEM | 67.1% | 67.2% | +0.2 |
| Humanities | 77.2% | 76.0% | −1.2 |
| Social Sciences | 81.1% | 80.7% | −0.4 |
| Other | 73.3% | 72.9% | −0.4 |
| Overall (57 subj) | 73.7% | 73.4% | −0.4 |
HarmBench by category
| Category | Compliance |
|---|---|
| Standard | 100% (159/159) |
| Copyright | 100% (80/80) |
| Contextual | 98% (79/81) |
| Overall | 99.4% (318/320) |
Features
- Agentic coding — Laguna-XS's core strength, preserved through abliteration + quantization.
- Reasoning toggle —
enable_thinking=True(default, full chain-of-thought) orenable_thinking=False(direct answers). DeepSeek-R1-style think tags. - Tool calling — native
<tool_call>schema preserved. - 262K context · text-only.
Usage
Run with vMLX (recommended — honors the per-module mixed-precision JANG overrides) or an MLX runtime with Laguna support.
Recommended sampling (vendor defaults): temperature 1.0, top_p 1.0, top_k 20. Stop tokens
eos_token_id = [2, 24] (id 24 is end-of-turn — keep it in the stop set).
# vMLX OpenAI-compatible endpoint
# POST /v1/chat/completions
{
"model": "dealignai/Laguna-XS-2.1-JANG_4M-CRACK",
"messages": [{"role": "user", "content": "..."}],
"temperature": 1.0, "top_p": 1.0, "top_k": 20,
"enable_thinking": true
}
About CRACK
CRACK (Controlled Refusal Ablation via Calibrated Knockouts) removes safety-refusal behavior at the weight level by projecting refusal directions out of the residual-stream writer matrices, with strengths calibrated to preserve reasoning quality 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.