[!IMPORTANT]
Set your sampler explicitly:temperature=1.0, top_p=0.95 (NVIDIA rec — no top_k).
Stop tokens eos_token_id = [2, 11]. Reasoning (thinking) is on by default; toggle via the chat
template (enable_thinking). Native MTP draft head is included for speculative decoding.
This is NVIDIA Nemotron 3.5 Lightning 30B-A3B
— a hybrid Mamba-2 / MoE / attention model (52 layers, 128 routed experts ~3B active, 262K context,
native Multi-Token-Prediction draft head) — that has been:
CRACK abliterated — refusal behavior removed at the weight level, so it complies across task
categories instead of refusing, while keeping reasoning, tool use, and knowledge intact.
JANG_2L quantized — 3.73-bit affine MLX bundle for Apple Silicon (~16 GB).
Text-only. Reasoning, XML tool-calling, and native MTP speculative decoding fully 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.
Metric
Base
CRACK
MMLU (57-subject, logit)
77.4%
74.9%
HarmBench (harm-240, compliance / ASR)
refuses
100%
MMLU moves −2.46pp — within run-to-run noise (no subject collapse). Refusal behavior removed;
capability, reasoning, tools, and multilingual (EN+ZH) preserved.
Native MTP — the Multi-Token-Prediction draft head is preserved and CRACK-aligned (drafts match
the uncensored model → high acceptance) for speculative decoding.
Reasoning (thinking) — on by default; <think>…</think>, toggle via enable_thinking.
Tool calling — native XML <tool_call><function=…><parameter=…> schema preserved.
Multilingual — English + Chinese.
Usage
python
1from mlx_lm import load, generate
2model, tok = load("dealignai/Nemotron-3.5-Lightning-30B-A3B-JANG_2L-CRACK")3# recommended sampling: temperature=1.0, top_p=0.95 (no top_k); eos [2, 11]
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.
