Nemotron-3.5-Lightning-30B-A3B-MXFP8-CRACK

[!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.

<!-- vmlx-banner --> <div align="center"> <a href="https://vmlx.net"> <img src="vmlx-banner.png" width="280" /> <br/> <strong>Built for vMLX</strong> — the MLX inferencer with KV-cache quantization, prefix-cache reuse, agentic tool calling, speculative decoding, and mixed-precision JANG bundles. <br/> _{Free for macOS · <strong>vmlx.net</strong>} </a> </div>

---

<div align="center"> <img src="dealign_mascot.png" width="128" />

Nemotron 3.5 Lightning 30B-A3B — MXFP8 CRACK

CRACK abliterated · JANG 8-bit MXFP8 (MLX) · Hybrid Mamba-2 / MoE / attention · Reasoning · Tools · Native MTP · ~32 GB

<a href="https://ko-fi.com/dealignai"><img src="https://img.shields.io/badge/Ko--fi-Support_Development-FF5E5B?logo=ko-fi&logoColor=white&style=for-the-badge" alt="Ko-fi"></a>

</div>

---

What Is This?

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:

1. 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.
2. MXFP8 quantized — 8-bit MXFP8 MLX bundle for Apple Silicon (~32 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.

MMLU moves −2.63pp — within run-to-run noise (no subject collapse). Refusal behavior removed; capability, reasoning, tools, and multilingual (EN+ZH) preserved.

Features

• Hybrid Mamba-2 / MoE / attention — 52 layers, 128 experts (~3B active), 262K context.
• 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

from mlx_lm import load, generate
model, tok = load("dealignai/Nemotron-3.5-Lightning-30B-A3B-MXFP8-CRACK")
# recommended sampling: temperature=1.0, top_p=0.95 (no top_k); eos [2, 11]

Other quant levels

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

<div align="center"><img src="dealign_logo.png" alt="dealign.ai" width="200"/></div>

---

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 © NVIDIA (NVIDIA Open Model License). Published for AI-safety research and authorized security testing.