Yoshi Suhara

The model was compressed from Nano 9B v2 using Nemotron Elastic (arxiv.org/abs/2511.16664) and post-trained on Nemotron 3 post-training data using a new recipe designed to improve accuracy in reasoning-off mode. Please see the blog and the Nemotron Elastic paper for details.

Quantized variants are also available! FP8 🤗 huggingface.co/nvidia/NVIDIA-… GGUF (Q4_K_M) 🤗huggingface.co/nvidia/NVIDIA-…

Nemotron 3 Nano 4B is here! A compact hybrid Mamba-Transformer model optimized for edge AI, with top-tier instruction following, strong tool use, and low-VRAM efficiency across NVIDIA Jetson, Spark, and RTX platforms. HF 🤗 huggingface.co/nvidia/NVIDIA-… Blog📝 huggingface.co/blog/nvidia/ne…

Introducing NVIDIA Nemotron 3 Super 🎉 Open 120B-parameter (12B active) hybrid Mamba-Transformer MoE model Native 1M-token context Built for compute-efficient, high-accuracy multi-agent applications Plus, fully open weights, datasets and recipes for easy customization and deployment. 🧵

【新モデル公開🚀Nemotron-Nano-9B-v2-Japanese】 本日、NVIDIA は Nejumi Leaderboard 4 のパラメータ数 10B 以下において、最先端の性能 (SOTA) を達成した NVIDIA Nemotron-Nano-9B-v2-Japanese を公開しました。商用利用可能です。 huggingface.co/blog/nvidia/ne…

🚀 New NVIDIA report: NVFP4 + Quantization-Aware Distillation (QAD) FP4 inference without quality collapse. Key idea: distill a BF16 teacher into an NVFP4 student using KL loss - much more robust than PTQ/QAT, especially after SFT/RL. 🔥 Near-BF16 accuracy ⚡ ~2-3× throughput, ~1.8× memory savings vs FP8 🧠 Works for LLMs and VLMs (Nemotron Nano, Super, VL) Technical report: huggingface.co/nvidia/NVIDIA-… Research blog: research.nvidia.com/labs/nemotron/… Hugging Face models: research.nvidia.com/labs/nemotron/…

🚀@Nvidia Nemotron 3 Nano is live! Nemotron 3 Nano is the world's most efficient open MoE with an Hybrid-MoE architecture and 1M context length. 🔥 Strong in reasoning, agentic and chat tasks with leading accuracy among AA index, Tau2, SWE Bench. 🔥 Up to 3.3X higher throughput comparing to other open MoE at similar sizes 🔥 A fully open recipe with data, infra released to the community Checkout the new model architecture and reinforcement learning technologies we used below: 😊 Huggingface: huggingface.co/collections/nv… 📢 Research blog: nvda.ws/48RusVt 🛣️Nemo RL & Nemo Gym (RL environment orchestration): github.com/NVIDIA-NeMo/RL & github.com/NVIDIA-NeMo/Gym Kudos to the teams for months of hard work! We are excited to keep building the Nemotron 3 model family and empower the community.

In collaboration with NVIDIA, the new Nemotron 3 Nano model is fully supported in llama.cpp Nemotron 3 Nano features an efficient hybrid, Mamba, MoE architecture. It's a promising model, suitable for local AI applications on mid-range hardware. The large context window makes it a great choice for a variety of use cases and applications. The efficiency of llama.cpp and the unique context management features of the `llama-server` tool allows us to deploy and use this model on a wide-range of hardware. With recent code contributions by engineering teams at NVIDIA and open-source collaborators, we can run this model very efficiently across the entire spectrum of NVIDIA GPUs. Learn more at @NVIDIA_AI_PC developer.nvidia.com/blog/inside-nv…

Today, @NVIDIA is launching the open Nemotron 3 model family, starting with Nano (30B-3A), which pushes the frontier of accuracy and inference efficiency with a novel hybrid SSM Mixture of Experts architecture. Super and Ultra are coming in the next few months.

I’m proud to be part of this amazing cross-team, cross-country collaboration supporting Sovereign AI for Japan 🇯🇵! HF 🤗: huggingface.co/datasets/nvidi… Blog 📝: huggingface.co/blog/nvidia/ne…

Thrilled to share that CLIMB has been accepted to the NeurIPS DB track! 🍀 Feeling so lucky to work with such an amazing team. #NeurIPS2025

NVIDIA has released Nemotron Nano 9B V2, a small 9B reasoning model that scores 43 on the Artificial Analysis Intelligence Index, the highest yet for <10B models Nemotron 9B V2 is the first Nemotron model pre-trained by @NVIDIA. Previous Nemotron models have been developed by post-training on Meta Llama models. Architecture & Training: The model uses a hybrid Mamba-Transformer architecture. NVIDIA pre-trained a 12B parameter base model and applied post-training with a range of techniques including RLHF and GRPO. The final 9B size was pruned from this model and re-trained with the base model as a teacher. Small-model frontier: with only 9B parameters, Nemotron Nano 9B V2 is placed ahead of Llama 4 Maverick on our leaderboard, equal to Solar Pro 2 with reasoning and trails just behind gpt-oss-20B (high). Along with this model, NVIDIA released a 6.6-trillion token subset of their pre-training data for public use on @huggingface Key model details: ➤ 128k token context window ➤ Supports reasoning and non-reasoning modes (with ‘/no_think’ settings in the system prompt) ➤ Released under the NVIDIA Open Model License, and not additionally covered by Meta’s Llama license like prior Nemotron models - this means that there is no limitation on use by large companies or requirement to keep ‘Nemotron’ in the name of derivative models ➤ No serverless inference providers are yet serving the model, but it is available now on Hugging Face for local inference or self-deployment See below for our full analysis and key announcement links from NVIDIA 👇

RT @shizhediao: ✨ Alongside NVIDIA-Nemotron-Nano-v2-9B, we’re also open-sourcing its pre-training dataset. At NVIDIA, we remain committed…

We're excited to share leaderboard-topping 🏆 NVIDIA Nemotron Nano 2, a groundbreaking 9B parameter open, multilingual reasoning model that's redefining efficiency in AI and earned the leading spot on the @ArtificialAnlys Intelligence Index leaderboard among open models within the same parameter range. It's built on a unique hybrid Transformer-Mamba architecture, a combination that delivers the same accuracy you expect, but with higher throughput. This enables it to achieve high performance/cost, making it perfect for real-world applications like customer service agents and chatbots. 🏗️ Hybrid Architecture: By combining the strengths of Transformer and Mamba architectures, achieves up to 6X faster throughput compared to other 8B open models and highest reasoning accuracy. 🏦 Thinking Budget: Reduces unnecessary token generation to cut costs by up to 60%, making it an ideal solution for balancing performance and total cost of ownership (TCO). 🔢 Open Datasets: The training datasets of this model are fully open, giving maximum transparency in using the model for enterprise applications. 🤗 Technical details on @HuggingFace ➡️ nvda.ws/3JfcKST 🏆 Leaderboard ➡️ nvda.ws/47B7iUh

We are excited to release Nvidia-Nemotron-Nano-V2 model! This is a 9B hybrid SSM model with open base model and training data. This model also supports runtime "thinking" budget control. HF collection with base and post trained models: huggingface.co/collections/nv…

📢New efficient Hybrid-SLM from NVIDIA-Nemotron-Nano-v2-9B: ❗️6x faster than Qwen3-8B because of Hybrid (Mamba2+Attention) design. We tried something new: pretrain & align a 12B reasoning model → compress to 9B. First real stab at reasoning-model compression. Key takeaways from compression: ▪️Target was 23GB GPUs + room for a 650M vision encoder → design via compression, not bespoke architecture. ▪️Distillation loss went down, but benchmarks didn’t - unlike Base compression. ▪️Reasoning compression needs light post-training alignment. ▪️Applied both Minitron + Puzzle. ▪️Dropped 2 attn layers to hit 128k context; KV cache dominated. ▪️Depth: 62 → 56 (fewer tanked accuracy). ▪️FFN: 20,480 → 15,680 (−23%). ▪️Hidden dimension: 5120 → 4480. ▪️Mamba heads → small gains <15%, mostly avoided. Distilled on 136B tokens, context grown 8k → 262k. It is required if we want to preserve long-context of 128k. 📰 Report: research.nvidia.com/labs/adlr/file… 🤗 HF: huggingface.co/collections/nv…

A new video game benchmark for LLM agents, designed across various game titles! Happy to be part of this wonderful collaboration with @dongmin_park11 and the amazing team @Krafton_AI!

Llama Nemotron Nano 4B matches the previous Nano 8B accuracy with only half the size! 🤗 huggingface.co/nvidia/Llama-3…

NeMo RL is now open source! It replaces NeMo-Aligner and is the toolkit we use to post train next generations of our models. Give it a try github.com/NVIDIA/NeMo-RL

Tool-using LLMs can learn to reason—without reasoning traces. 🔥 We present Nemotron-Research-Tool-N1, a family of tool-using reasoning LLMs trained entirely via rule-based reinforcement learning—no reasoning supervision, no distillation. 📄 Paper: arxiv.org/pdf/2505.00024 💻 Code: github.com/NVlabs/Tool-N1 (Please consider giving us a ⭐️ to stay updated on the upcoming code release!) 🧠 Why this matters: Existing tool-call models rely heavily on supervised reasoning traces from stronger models—costly, brittle, and often imitative. We ask: Can LLMs learn to reason directly from tool success signals? 📦 What we did: – Train Qwen2.5-7B/14B with simple binary reward on tool-call correctness + reasoning format in R1-style – No reasoning traces needed – Evaluate on BFCL, API-Bank, and ACEBench – Also study the role of SFT, RL, and widely adopted SFT-then-RL recipes in training Tool-Calling models. 📈 Key findings: – Tool-N1-7B/14B obviously outperform GPT-4o and open baselines on all benchmarks – Widely adopted SFT+RL paradigm doesn’t necessarily lead to better performance than Pure RL. – Binary reward > fine-grained reward, esp. for real-world queries – Scaling works: bigger = better gains under our RL setup 🌟 Takeaway: Reasoning doesn’t have to be taught. With just a binary signal, LLMs can learn to reason and act. Tool-N1 sets a new direction for scalable, supervision-light tool calling model training