openNemo-9B-Claude-Opus-4.6-distill

Reasoning-distilled version of openNemo-9B, fine-tuned on Claude Opus 4.6 reasoning traces.

Trained with SFT + DPO on community-curated reasoning distillation datasets to produce step-by-step <think> chains before answering. Built on the openNemo pure-PyTorch Nemotron-H architecture — no mamba-ssm or causal-conv1d required.

By Empero AI

What is this?

A 9B dense hybrid model (Mamba2 + Transformer) that has been taught to reason through problems before answering, using reasoning traces distilled from Claude Opus 4.6 and other frontier models.

The two-stage training pipeline:

1. SFT — teaches the reasoning format: <think> tags, step-by-step chains, edge-case consideration
2. DPO — teaches preference for thorough reasoning over skipping the thinking step

Quickstart

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

# 4-bit quantization (fits in ~8 GB VRAM)
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

model = AutoModelForCausalLM.from_pretrained(
    "empero-ai/openNemo-9B-Claude-Opus-4.6-distill",
    quantization_config=bnb_config,
    trust_remote_code=True,
    device_map="auto",
)

tokenizer = AutoTokenizer.from_pretrained("empero-ai/openNemo-9B-Claude-Opus-4.6-distill")

messages = [
    {"role": "system", "content": "You are a deep reasoning AI. When given a problem, you think through it carefully and methodically inside <think> tags before providing your final answer."},
    {"role": "user", "content": "Prove that the sum of the first n odd numbers equals n²."},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=2048, do_sample=True, temperature=0.7, top_p=0.95)
response = tokenizer.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(response)

Without thinking (instruct mode)

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)

Architecture

9B dense hybrid Nemotron-H — same architecture as the base openNemo-9B:

Training Details

Stage 1: Supervised Fine-Tuning (SFT)

Trained on 8 reasoning distillation datasets:

Stage 2: Direct Preference Optimization (DPO)

Preference pairs constructed from the same datasets:

• Chosen: Full response with <think> reasoning chain
• Rejected: Same response with <think> block stripped

Additional DPO source: QuietImpostor/Sao10K-Claude-3-Opus-Instruct-15K-ShareGPT

Hyperparameters

GGUF

Quantized GGUF versions are available at empero-ai/openNemo-9B-Claude-Opus-4.6-distill-GGUF.

Requirements

torch>=2.1
transformers>=4.40
bitsandbytes>=0.43  # for 4-bit quantization

No mamba-ssm. No causal-conv1d. No CUDA kernel compilation.

Base Model

This model is built on empero-ai/openNemo-9B, a pure-PyTorch drop-in replacement for NVIDIA's Nemotron-H that removes all external CUDA kernel dependencies. See the base model card for details on the architecture changes.

Citation

@misc{openNemo-9B-Claude-Opus-distill,
  title={openNemo-9B-Claude-Opus-4.6-distill},
  author={Empero AI},
  year={2026},
  url={https://huggingface.co/empero-ai/openNemo-9B-Claude-Opus-4.6-distill}
}

License

NVIDIA Open Model License — same as the base model.

Acknowledgments

• Base model: openNemo-9B by Empero AI
• Original architecture: Nemotron-H by NVIDIA
• Reasoning datasets: Community contributors (nohurry, Roman1111111, Crownelius, TeichAI, Jackrong, dalisoft, Hastagaras, Sao10K)