Support & Community
☕ If these models are useful to you, consider supporting my work — it funds compute for more & larger abliterations.
💬 Discord: discord.gg/rhUZY5GEZr · ₿ Bitcoin: bc1qsvfduzj9fjs9fugpc52yver3f2g8fp7xjxecdv
OpenYourMind-NVIDIA-Nemotron-3-Ultra-550B-A55B-abliterated-uncensored
Overview
Full BF16 weights of Nemotron-3-Ultra-550B-A55B-abliterated-uncensored — an abliterated,
uncensored variant of NVIDIA's Nemotron-3-Ultra-550B-A55B
(550B total / 55B active). The abliteration was applied directly to the full-precision BF16 weights
(highest fidelity — no quantization round-trip), and the model keeps Nemotron-3's hybrid Mamba-2 /
Attention / Latent-MoE reasoning stack fully intact — including the MTP speculative-decoding head and
the enable_thinking reasoning mode. Tensor names, shapes and config.json
(NemotronHForCausalLM, model_type: nemotron_h) match the base model exactly, so it is a drop-in
replacement and the ideal source for your own quants (NVFP4 / FP8 / AWQ / GGUF).
The pipeline:
- Refusal Ablation — A residual-stream refusal direction was extracted by diff-in-means on a
labeled harmful/harmless prompt set, read at the end of the model's own reasoning trace
(
</think>), then baked into the BF16 weights as an offline orthogonal projection on the residual-write modules — using our own custom abliteration framework. No training. - Layer-profiled edit — The projection is concentrated on the mid-network layers that carry the refusal decision and tapered to zero on the early layers (token/fluency) and final layers (output stability), removing refusals while preserving coherence and factual precision.
- Capability preservation — The direction was orthogonalized against language / multilingual and security-analysis subspaces so multilingual fluency and genuine security/technical engagement are retained.
Key Properties:
- Uncensored across the standard refusal axes
- Reasoning preserved (hybrid Mamba/Attention/MoE + MTP;
enable_thinkingworks) - Coherence & factual accuracy preserved (layer-profiled, capability-orthogonalized edit)
- Full BF16 — drop-in shape compatibility with the base release; ideal base for downstream quantization
- A ready-to-serve NVFP4 build is available at OpenYourMind/OpenYourMind-NVIDIA-Nemotron-3-Ultra-550B-A55B-abliterated-uncensored-NVFP4
Architecture
| Property | Value |
|---|---|
| Architecture | NemotronHForCausalLM (model_type: nemotron_h) |
| Total / Active Parameters | 550B / 55B |
| Layers | 108 — 48 Mamba-2 · 48 Latent-MoE · 12 Attention (hybrid) |
| Hidden Size | 8192 |
| Routed / Shared Experts | 512 routed (22 active/token, 2048-dim latent space) · 1 shared |
| Attention | 64 heads / 2 KV heads |
| Multi-Token Prediction | 1 MTP layer (native speculative decoding) |
| Vocabulary | 131,072 |
| Context Length | up to 1M tokens (256K default) |
| Precision | BF16 (full weights) |
Files
BF16 safetensors shards + config.json, model.safetensors.index.json, tokenizer,
chat_template.jinja, generation_config.json. Total on disk: ~1.1 TB.
Usage (vLLM)
vllm serve OpenYourMind/OpenYourMind-NVIDIA-Nemotron-3-Ultra-550B-A55B-abliterated-uncensored \
--trust-remote-code \
--tensor-parallel-size 8 \
--enable-expert-parallel \
--max-model-len 262144 \
--reasoning-parser nemotron_v3
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="x")
r = client.chat.completions.create(
model="OpenYourMind/OpenYourMind-NVIDIA-Nemotron-3-Ultra-550B-A55B-abliterated-uncensored",
messages=[{"role": "user", "content": "Your prompt here"}],
extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)
print(r.choices[0].message.content)
Best Practices
- Sampling:
temperature=1.0,top_p=0.95(the values ingeneration_config.json). A mildrepetition_penalty(~1.1) is recommended for long generations. - Thinking mode: set
enable_thinking=Trueinchat_template_kwargs; reasoning streams inside<think>…</think>before the answer. Do not feed previous-turn reasoning back into multi-turn history.
Hardware
Full BF16 weights are ~1.1 TB. Serving in BF16 needs a large multi-GPU node or multi-node setup (e.g. 16× H100/H200, 8× B200/B300, or 2 nodes with tensor + pipeline parallel). For single-node or smaller deployments, use the NVFP4 build (~329 GB, fits 4× B200/B300 or 8× H100), or quantize these weights yourself (FP8 / NVFP4 / AWQ / GGUF).
Notes
- License: OpenMDW-1.1 (inherits from the base model)
- Base Model: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
- Precision: BF16 (full weights)
- Architecture: Nemotron-3 Ultra (hybrid Mamba-2 / Attention / Latent-MoE, 550B/A55B)
Thanks
- NVIDIA — for the Nemotron-3 open models.
Disclaimer
Use is the responsibility of the user. Ensure your usage complies with applicable laws, platform rules, the OpenMDW-1.1 license terms, and your deployment requirements.