Wonmin Byeon

🚀 New paper: Mamba–Transformer hybrid VLMs can go fast without forgetting. We introduce stateful token reduction for long-video VLMs. ✅ Only 25% of visual tokens 🚀 3.8–4.2× faster prefilling (TTFT) 🎯 Near-baseline accuracy (can exceed baseline with light finetuning)

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. 🧵

Announcing NVIDIA Nemotron 3 Super! 💚120B-12A Hybrid SSM Latent MoE, designed for Blackwell 💚36 on AAIndex v4 💚up to 2.2X faster than GPT-OSS-120B in FP4 💚Open data, open recipe, open weights Models, Tech report, etc. here: research.nvidia.com/labs/nemotron/… And yes, Ultra is coming!

One of the things that makes Nemotron 3 Super so fast is native multi-token prediction. 1. Model predicts several tokens rather than just one, which is essentially free because it's just a bit of extra work for the last layer of the model. The first token is accepted, the others are provisional. 2. On the next pass, the model verifies the extra tokens and either accepts them if they were correct or rejects them if they were wrong. If they were wrong, we still accept one token, so autoregression proceeds even with misprediction. 3. Verification is essentially free because the GPU has a lot of extra compute capability when running at a small batch size that would otherwise be wasted. 4. Speedup then is directly linked to acceptance rates. If the multi-token prediction is more accurate, we get a bigger speedup.

@bronzeagepapi It can be used for gated deltanet as well. Hopefully, someone tries it with Qwen3.5 VL ;)

@wonmin_byeon why mamba over gated deltanet?

🚀 New paper: Mamba–Transformer hybrid VLMs can go fast without forgetting. We introduce stateful token reduction for long-video VLMs. ✅ Only 25% of visual tokens 🚀 3.8–4.2× faster prefilling (TTFT) 🎯 Near-baseline accuracy (can exceed baseline with light finetuning)

5) Results: with only 25% visual tokens, we get 3.8–4.2× faster prefilling (TTFT) while keeping near-baseline accuracy. With light finetuning under reduction, accuracy improves further and can even surpass baseline.

4) Key piece: language-aware token scoring across attention and SSM: - attention layers (explicit attention) - Mamba/SSM layers (implicit-attention proxy) So selection stays prompt-aligned.

@Jeongseok_hyun @NVIDIAAI @RHachiuma Welcome to NVIDIA!

🎉Thrilled to share that I will be joining NVIDIA Taiwan (@NVIDIAAI) as a research intern! I'll be working on efficient modeling of Video LLMs under the supervision of Ryo Hachiuma (@RHachiuma). Looking forward to this exciting journey ahead!

AI Coffee Break with Letitia: youtu.be/uMk3VN4S8TQ?si… Arxiv Papers: youtu.be/I_loecu6Gew?si… AI Papers Podcast: youtu.be/ifqHBtpw7CM?si… AI Paper Slop: youtu.be/pKObqY3msJk?si… Our project page and the full paper are here: research.nvidia.com/labs/lpr/storm/

It’s been months since we introduced STORM, and it’s great to see so many video discussions! ⛈️🎥 To catch up on our token-efficient long video understanding work, check out these reviews (Some are AI-generated but very useful!). Huge thanks to the creators! 🙌 The links ↓

What if your model could learn from its own drafts during RL training? 🚀 🔥 New paper: iGRPO: Iterative Group Relative Policy Optimization We add a self-feedback loop to GRPO: the model drafts multiple solutions, picks its best one, then learns to refine beyond it. Core idea: Stage 1 → explore and select your strongest attempt. Stage 2 → condition on that attempt and beat it. Same scalar reward. No critics, no generated critiques, no verification text. The best draft is the only feedback the model needs. 📊 Results across 7B / 8B / 14B models: • Nemotron-H-8B-Base-8K: 41.1% → 45.0% (+3.96 over GRPO) • DeepSeek-R1-Distill-Qwen-7B: 68.3% → 69.9% • OpenMath-Nemotron-14B: 76.7% → 78.0% • OpenReasoning-Nemotron-7B on AceReason-Math: 85.62% AIME24 / 79.64% AIME25 🏆 The same two-stage wrapper also improves DAPO and GSPO. It's not tied to GRPO at all. 📄 Paper: arxiv.org/abs/2602.09000 This work was the result of a wonderful collaboration with @shrimai_ @igtmn @GXiming @SeungjuHan3 @_weiping @YejinChoinka @jankautz

🚀Announcing C-RADIOv4: The latest evolution in our agglomerative vision backbone family is here! We’ve built a unified student model that distills the best capabilities of multiple state-of-the-art teachers into a single, efficient architecture. DINOv3, SAM3 and SigLIP2 all in one forward pass with better features. Better efficiency, robustness and any resolution. 👐Permissive and open-source! Read the tech report: arxiv.org/pdf/2601.17237 Models: huggingface.co/nvidia/C-RADIO… huggingface.co/nvidia/C-RADIO… Github: github.com/NVlabs/RADIO 🧵1/N

Nvidia is fast becoming open-source kings, they actually pre-trained, mid-trained, post-trained, and got better results than Qwen3 which is impressive! Now we only need to bully them about nvidia-open-model-license and multimodality...

🚀 Nemotron-3 Nano is out. An open-weight reasoning model built for efficiency, long-context intelligence, and agentic workflows. Faster inference. Stronger reasoning. Built to scale. 📄 Technical report ↓ research.nvidia.com/labs/nemotron/… #Nemotron3 #AI #OpenModels #NVIDIA

We just released Nemotron 3 Nano 30B-A3B. It is fast, accurate, and truly open. Check it out! Nemotron 3 Super (~4x larger) and Ultra (~16x larger) are currently being trained and will be released in 2026! blog-post and technical report: nvda.ws/48RusVt

NVIDIA’s hybrid model just got bigger, faster, and stronger.