Fuxiao Liu

Research Scientist @Nvidia | CS PhD @UMDCSI, working on LLM, Multimodal Stuff

[Furong Huang @furongh]

While FIFA has captured everyone’s attention, something consequential happened in academia this week, especially for international students in PhD programs.

DHS issued a final rule changing how long F-1 students may remain in the United States.

A thread on what it means and what students and universities can do. 🧵

[Yunze Man @yunzeman]

> Generation can become unstable if the harness drops thinking history or switches models mid-session

For long context agents, thinking history is becoming part of the harness. Model release is never just a model anymore, they ship with a harness

I feel this with GPT5.6-Sol. In web it feels no big deal, in codex it feels like cheating

[Shilong Liu @Shilong_Liu_AI]

https://t.co/lkoQPTFVnq

[Alisa Liu @alisawuffles]

I'm joining OpenAI next week!🥹 The job search turned out to be really challenging but also super rewarding, so I wrote a small blog to share what I learned along the way and hopefully make the process a little less mysterious for the next person. https://t.co/6FigSBdenD

[NVIDIA AI @NVIDIAAI]

Today we're shipping Nemotron 3 Ultra.

A 550B MoE frontier-intelligence open model built for long-running agents.

It delivers 5x faster inference and lowers the cost of complex agentic tasks by up to 30% versus other open frontier models.

[Fuxiao Liu @Kongkong]

Try our Nemotron 3 Nano Omni Demo at: https://t.co/KzNeQGAks6

[AK @_akhaliq]

Nvidia released Nemotron 3 Nano Omni

made a gradio app for it on Hugging Face

[AK @_akhaliq]

Nvidia released Nemotron 3 Nano Omni

made a gradio app for it on Hugging Face

[Fuxiao Liu @Kongkong]

Stop choosing between efficiency and accuracy for your AI agents. 🛠️ NVIDIA Nemotron 3 Nano Omni is here:
✅ Unified Reasoning: One model for video, audio, and text.
✅ Up To 9x Throughput: Massive efficiency gains for video workflows.
✅ Fully Open: Weights + Data + Recipes. Optimized for Blackwell and available as a NIM. Let’s build. 🚀

Hugging Face: https://t.co/jwGRGbfLKG
Tech Report: https://t.co/4FTsBVE5bq

#nvidia #nemotron #omni

[Fuxiao Liu @Kongkong]

Stop choosing between efficiency and accuracy for your AI agents. 🛠️ NVIDIA Nemotron 3 Nano Omni is here:
✅ Unified Reasoning: One model for video, audio, and text.
✅ Up To 9x Throughput: Massive efficiency gains for video workflows.
✅ Fully Open: Weights + Data + Recipes. Optimized for Blackwell and available as a NIM. Let’s build. 🚀

Hugging Face: https://t.co/jwGRGbfLKG
Tech Report: https://t.co/4FTsBVE5bq

#nvidia #nemotron #omni

[NVIDIA @nvidia]

https://t.co/Ow621ET16t

[Bryan Catanzaro @ctnzr]

Thank you to everyone in the community who is testing and using Nemotron models. It's great to see Nemotron-Cascade-2, Nemotron-3-Super and Nemotron-3-Nano trending on HF.

The Nemotron team is working hard to incorporate all your feedback into Nemotron 4.

And yes, Nemotron 3 Ultra is still on track for release.

https://t.co/lkEwmlUng9

[Fuxiao Liu @Kongkong]

What if a VLM could teach itself from zero data?

Meet MM-Zero: one base model that self-evolves by playing 3 roles:
🧠 Proposer – imagines visual concepts
💻 Coder – renders them in Python/SVG
🔎 Solver – reasons over the results

No seed images. Just RL (GRPO) + execution feedback. @zli12321 @ChengsongH31219

[AK @_akhaliq]

MM-Zero

Self-Evolving Multi-Model Vision Language Models From Zero Data

paper: https://t.co/o5d40EF8yo

[Fuxiao Liu @Kongkong]

🧠 NVIDIA Nemotron 3 Super is the best open-source brain for OpenClaw.

#nvidia #openclaw #clawdbot #nemotron

[Bryan Catanzaro @ctnzr]

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:
https://t.co/CAYpP1iK3i

And yes, Ultra is coming!

[Bryan Catanzaro @ctnzr]

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:
https://t.co/CAYpP1iK3i

And yes, Ultra is coming!

[Fuxiao Liu @Kongkong]

🔥🔥🔥

[Bryan Catanzaro @ctnzr]

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:
https://t.co/CAYpP1iK3i

And yes, Ultra is coming!

[ChengSong Huang @ChengsongH31219]

We remove the requirement of image during the vlm training in MM-Zero!!!

[Guilherme Favaron @guifav]

What if a vision model could learn to reason about images without ever seeing one?

MM Zero, from Zongxia Li, @kongkong, and researchers at University of Maryland, Brown, Adobe, and NVIDIA, introduces a three role framework where VLMs bootstrap visual reasoning from literally zero data.

Three agents (Proposer, Coder, Solver) all start from the same base model. The Proposer invents visual concepts and questions. The Coder renders them as SVG code. The Solver reasons over the results. All trained via GRPO reinforcement learning, no human annotation.

On visual reasoning benchmarks, Qwen3 VL 8B improves from 50.7% to 56.6% accuracy. Mimo VL 7B goes from 50.9% to 56.0%. Performance keeps climbing through 5 iterations with no sign of plateau.

The data bottleneck for VLM training may be less about collecting more images and more about letting models generate their own.

[Fuxiao Liu @Kongkong]

@simon_zhai All the best