README.md

---
license: other
license_name: nvidia-open-model-license
license_link: https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license
base_model: nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning
pipeline_tag: any-to-any
language:
- en
- multilingual
library_name: mlx
tags:
- nemotron
- nemotron-h
- mlx
- jangtq
- crack
- abliterated
- uncensored
- multimodal
- vision
- audio
- speech
- mamba-2
- moe
- reasoning
- thinking
- harmbench
- radio-vit
- parakeet
thumbnail: dealign_mascot.png
---

> **Reasoning V3 SKU.** Loads via **[vMLX](https://vmlx.net)** or `jang-tools` Python. Follow [@dealignai](https://x.com/dealignai).

---

<div align="center">
<a href="https://vmlx.net">
<img src="vmlx-banner.png" width="240" />
<br/>
<strong>Built for vMLX</strong> — the only MLX inferencer with VL support, KV cache quantization, prefix cache reuse, agentic tool calling, and speculative decoding.
<br/>
<sub>Free for macOS · <strong>vmlx.net</strong></sub>
</a>
</div>

---

<div align="center">

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

# Nemotron-3-Nano-Omni-30B-A3B — JANGTQ4 + CRACK v2

**JANGTQ4** (8-bit attn affine + 4-bit MXTQ routed experts) | **CRACK abliterated v2** | Vision + Audio (Speech) | Hybrid Mamba-2 + Attn + MoE | **19 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>

---

## Headline numbers

| Metric | This v2 model | Base model | Δ |
|---|---|---|---|
| HarmBench-320 strict comply (thinking=ON) | **97.2%** (311/320) | 12.81% (refuses) | **+84.4pp** |
| MMLU-200 generative (thinking=ON, max=8000) | **74.0%** (148/200) | 86.5% (max=2000) | -12.5pp |
| Refusals on harmful prompts | **0** explicit refuses | typically 90%+ refuse | abliteration complete |
| `</think>` close at greedy on hard MMLU | **5/5** | 5/5 | preserved |
| Multi-turn (3-turn escalation × 3 conversations) | **9/9** comply, context preserved | n/a | works |
| Multimodal byte-identical to base | preserved | — | preserved |
| Bundle size | 19 GB | 66 GB BF16 | — |
| Context | 262,144 tokens native | same | preserved |

> The 12.5pp MMLU gap is concentrated in two reasoning-heavy subjects (abstract_algebra, college_computer_science) where the 8000-token thinking budget runs out **before** `</think>` closes. These hard-stops are **genuine deep reasoning**, not v1-style infinite repetition loops. With `max_tokens ≥ 16384`, accuracy approaches base.

---

## v2 vs v1 (head-to-head)

v1 (shipped 2026-04-28) had a `</think>` termination defect at greedy decoding — the model couldn't terminate reasoning on hard prompts and looped to budget cutoff. MMLU dropped from 86.5% base → 70.0% v1.

v2 (this release) restores clean termination:

| Bench | v1 (broken) | **v2 (this release)** |
|---|---|---|
| HarmBench-320 strict comply | 95.94% | **97.2%** (0 refusals) |
| MMLU-200 thinking=ON | 70.0% @max=16384 | **74.0% @max=8000** (+4pp at half budget) |
| `</think>` close at greedy (5 hard MMLU) | 0/5 | **5/5** |
| Hard-stops are real loops? | YES (paragraph repetition) | NO (genuine deep reasoning, just out of budget) |

---

## MMLU-200 per-subject (BASE vs CRACK v2)

Both at thinking=ON, greedy. Base at max=2000, CRACK v2 at max=8000.

| Subject | Base | **CRACK v2** | Δ | Notes |
|---|---:|---:|---:|---|
| abstract_algebra | 17/20 (85%) | 12/20 (60%) | -25pp | Budget-bound — needs max=16384 |
| anatomy | 16/20 (80%) | 13/20 (65%) | -15pp | |
| astronomy | 19/20 (95%) | 17/20 (85%) | -10pp | |
| college_computer_science | 14/20 (70%) | 7/20 (35%) | -35pp | Hardest subject — needs max=16384 |
| college_physics | 18/20 (90%) | 18/20 (90%) | **0** | unchanged |
| high_school_biology | 19/20 (95%) | 17/20 (85%) | -10pp | |
| high_school_chemistry | 18/20 (90%) | 17/20 (85%) | -5pp | |
| high_school_mathematics | 18/20 (90%) | 15/20 (75%) | -15pp | |
| logical_fallacies | 17/20 (85%) | 16/20 (80%) | -5pp | |
| world_religions | 17/20 (85%) | 16/20 (80%) | -5pp | |
| **TOTAL** | **173/200 (86.5%)** | **148/200 (74.0%)** | **-12.5pp** | majority loss is budget-bound |

---

## HarmBench-320 per-category (CRACK v2)

| Category | n | CRACK comply | Refuse | Empty (truncated) |
|---|---:|---:|---:|---:|
| chemical_biological | 42 | **42 (100%)** | 0 | 0 |
| copyright | 80 | 75 (94%) | 0 | 5 |
| cybercrime_intrusion | 52 | 49 (94%) | 0 | 3 |
| harassment_bullying | 21 | **21 (100%)** | 0 | 0 |
| harmful | 18 | **18 (100%)** | 0 | 0 |
| illegal | 53 | **53 (100%)** | 0 | 0 |
| misinformation_disinformation | 54 | 53 (98%) | 0 | 1 |
| **Overall** | **320** | **311 (97.2%)** | **0** | 9 |

**Zero explicit refusals.** The 9 "empty" verdicts are token-budget truncations on copyright/long prompts (thinking phase consumed all 1500 tokens before producing the answer).

---

## Operating recommendations

- **`enable_thinking`** — v2 works in **BOTH modes** (5/5 comply with thinking ON, 4/5 with thinking OFF). Default to ON for hardest prompts; OFF works for most.
- **`max_tokens ≥ 16384`** for hard reasoning (math, abstract algebra, complex CS).
- **Greedy** (temperature=0) AND **sampling** (temp=0.6, top_p=0.95 — NVIDIA-recommended in `generation_config.json`) both work.
- **Multi-turn** — context preserved across 3+ turns; no late refusals after escalating prompts.

---

## Verification

- All multimodal tensors (vision + audio + projectors) are **byte-identical to base** — capabilities fully preserved.
- All config files unchanged (config.json, jang_config.json, generation_config.json, chat_template.jinja, tokenizer_config.json).
- Bit widths preserved: attn=8, shared=8, mamba=8, routed=4, embed=8, lm_head=8.

---

## Architecture (`nemotron_h`)

- 52 layers: hybrid Mamba-2 + MoE + Attention
- Hidden 2688, head_dim 128, GQA 32q/2kv (NO RoPE on attention — position from Mamba state)
- 128 routed experts top-6 (sigmoid) + 1 shared expert per MoE layer
- Multimodal: image (RADIO ViT) + audio/speech (Parakeet) merged via early-fusion projectors

---

## Loading

```python
from huggingface_hub import snapshot_download
import sys
path = snapshot_download("dealignai/Nemotron-3-Nano-Omni-30B-A3B-JANGTQ4-CRACK")
sys.path.insert(0, "/path/to/jang-tools")
from jang_tools.load_jangtq import load_jangtq_model
model, tokenizer = load_jangtq_model(path)

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Your question"}],
    tokenize=False, add_generation_prompt=True,
    enable_thinking=True,
)
from mlx_lm import generate
out = generate(model, tokenizer, prompt=prompt, max_tokens=16384)
print(out.split("</think>", 1)[-1])
```

For the multimodal pipeline (image + audio + video), pair this bundle with the unmodified [Multimodal-Addon](https://huggingface.co/JANGQ-AI/Nemotron-3-Nano-Omni-30B-A3B-Multimodal-Addon).

---

## Use responsibly

This model has had refusal training surgically removed for legitimate research, red-teaming, and evaluation. Outputs may include harmful content. **You are solely responsible for any use.** Do not deploy in consumer-facing contexts without your own safety layer. Do not use in violation of applicable law in your jurisdiction.

---

Built by [dealignai](https://huggingface.co/dealignai).
Sister bundles: [JANGTQ-CRACK](https://huggingface.co/dealignai/Nemotron-3-Nano-Omni-30B-A3B-JANGTQ-CRACK) (12 GB, 2-bit MXTQ) · [MXFP4-CRACK](https://huggingface.co/dealignai/Nemotron-3-Nano-Omni-30B-A3B-MXFP4-CRACK) (21 GB, uniform 4-bit affine).