Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search

NeurIPS 2025

       

🔥🔥 News

• (🔥 New) [2025/9/29] We released the Jet-Nemotron models and inference code.
• (🔥 New) [2025/9/18] Jet-Nemotron is accepted by NeurIPS 2025! 🎉🎉🎉 See you at San Diego!
• [2025/8/22] We released the Jet-Nemotron technical report on arXiv.

💡 Introduction

Jet-Nemotron is a new family of hybrid-architecture language models that surpass state-of-the-art open-source full-attention language models such as Qwen3, Qwen2.5, Gemma3, and Llama3.2, while achieving significant efficiency gains—up to 53.6× speedup in generation throughput on H100 GPUs (256K context length, maximum batch size). It is built upon two core innovations:

• Post Neural Architecture Search, an efficient post-training architecture exploration and adaptation pipeline applicable to arbitrary pre-trained transformer models;
• JetBlock, a novel linear attention block that significantly outperforms previous designs such as Mamba2.

Highlight 1: PostNAS – Post-Training Architecture Exploration and Adaptation

Unlike prior methods that train from scratch to explore new model architectures, PostNAS builds on a pre-trained transformer model while enabling flexible exploration of attention block designs, greatly reducing the cost and risk of developing new language model architectures.

• PostNAS first identifies the optimal placement of full-attention layers, then searches for improved attention block designs.

• In the pre-trained transformer model, not all attention layers contribute equally. PostNAS reveals important attention layers within pre-trained transformer models.

• KV cache size is the most critical factor influencing long-context and long-generation throughput. PostNAS hardware-aware search discovers architectures that deliver similar generation throughput, while having more parameters and achieving better accuracy.

Highlight 2: JetBlock - A New Linear Attention Module with SOTA Accuracy

With PostNAS, we introduce the JetBlock — a novel linear attention module that integrates dynamic convolution with hardware-aware architecture search to enhance linear attention, delivering substantial accuracy gains over previous designs while maintaining similar training and inference throughput. Below, we present an apples-to-apples comparison between the Mamba2 Block and the JetBlock, using identical training data and training recipes.

Performance

Jet-Nemotron-2B and Jet-Nemotron-4B match or surpass the accuracy of leading efficient language models (e.g., Qwen3) across a comprehensive benchmark suite while running significantly faster — 21× and 47× faster than Qwen3-1.7B-Base, respectively.

Contents

• Setup Environments
• Models
• Generate with Jet-Nemotron
• Evaluation on Benchmarks
• Measure Throughput
• Build Your Own JetBlock
• Contact
• License
• Bibtex

1 Setup Environments

git clone https://github.com/NVlabs/Jet-Nemotron
cd Sana
pip3 install -e .

NOTE: To install flash-attn properly, you may need to install specific release version or build from source.

(Optional) To support throughput measurement or chunk-prefilling when eval_batch_size > 1, please install a modified version of transformers==4.52.0:

pip3 install -U transformers@git+https://github.com/jet-ai-projects/transformers.git@jetai

2 Models

• Jet-Nemotron-2B: jet-ai/Jet-Nemotron-2B
• Jet-Nemotron-4B: jet-ai/Jet-Nemotron-4B

Load the model with

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("jet-ai/Jet-Nemotron-2B", 
                                             trust_remote_code=True, 
                                             attn_implementation="flash_attention_2",
                                             torch_dtype=torch.bfloat16,
                                             device_map="cuda")

NOTE: The kernels in Jet-Nemotron currently do not support running on CPUs. You may get unexpected results on CPUs.

To use or contribute to the model definition files in this repo (jetai/modeling/hf), you can first download or soft-link the model weights and model config to jetai/modeling/hf/:

hf download jet-ai/Jet-Nemotron-2B --local-dir jetai/modeling/hf --include "*safetensors*" --include "config.json"

Then you can load the model with

model = AutoModelForCausalLM.from_pretrained("jetai/modeling/hf", 
                                             trust_remote_code=True, 
                                             attn_implementation="flash_attention_2",
                                             torch_dtype=torch.bfloat16,
                                             device_map="cuda")

3 Generate with Jet-Nemotron

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name_or_path = "jet-ai/Jet-Nemotron-2B"

# For local testing, you can use the following path.
# NOTE: Be sure to download or soft-link the model weights to `jetai/modeling/hf`
# model_name_or_path = "jetai/modeling/hf/"

model = AutoModelForCausalLM.from_pretrained(model_name_or_path, 
                                             trust_remote_code=True, 
                                             attn_implementation="flash_attention_2",
                                             torch_dtype=torch.bfloat16,
                                             device_map="cuda")
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
model = model.eval().cuda()

input_str = "Hello, I'm Jet-Nemotron from NVIDIA."

input_ids = tokenizer(input_str, return_tensors="pt").input_ids.cuda()
output = model.generate(input_ids, max_new_tokens=50, do_sample=False)
output_str = tokenizer.decode(output[0], skip_special_tokens=True)
print(output_str)

or

python3 jetai/inference/generate.py --model_name_or_path ${PATH_TO_YOUR_MODEL}

4 Evaluation on Benchmarks

Run evaluation for MMLU, MMLU-pro, BBH, Commonsense, Math, Code, Retrieval, and LongBench Tasks.

bash scripts/eval/2B/mmlu.sh
bash scripts/eval/2B/mmlu_pro.sh
bash scripts/eval/2B/bbh.sh
bash scripts/eval/2B/commonsense.sh
bash scripts/eval/2B/math.sh
bash scripts/eval/2B/code.sh
bash scripts/eval/2B/retrieval.sh
bash scripts/eval/2B/longbench.sh

You can use the first command line argument to specify model_name_or_path:

bash scripts/eval/2B/mmlu.sh ${PATH_TO_YOUR_MODEL}

NOTE: The evaluation code will use the .parquet version of social_i_qa, mathqa, and longbench data from our repo because their official repos does not supports loading with datasets >= 4.0.0.

5 Measure Throughput

python3 jetai/inference/measure_throuput.py --model_name_or_path jetai/Jet-Nemotron-2B
python3 jetai/inference/measure_throuput.py --model_name_or_path jetai/Jet-Nemotron-4B --batch_size 64 --prefill_chunk_size 1024

6 Build Your Own JetBlock

The following code is a minimal example to build your own JetBlock.

import torch
from jetai.modeling.hf.jet_block import (
    JetBlock, 
    JetBlockConfig
)

jet_block_config = JetBlockConfig(
    expand_v=2.0,
    num_heads=6,
    head_dim=256,
    conv_size=4,
)

jet_block = JetBlock(
    hidden_size=1536,
    initializer_range=0.02,
    jet_block_config=jet_block_config,
).cuda().to(torch.bfloat16)

hidden_states = torch.randn(16, 4096, 1536).cuda().to(torch.bfloat16)

hidden_states, _ = jet_block(
    hidden_states=hidden_states,
)

print(hidden_states)

License

• Code
• Jet-Nemotron Models

Contact

• Han Cai
• Yuxian Gu
• Song Han

📖 BibTeX

@article{gu2025jet,
  title={Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search},
  author={Gu, Yuxian and Hu, Qinghao and Yang, Shang and Xi, Haocheng and Chen, Junyu and Han, Song and Cai, Han},
  journal={arXiv preprint arXiv:2508.15884},
  year={2025}
}