Twitxy Saved

Ryan Hanrui Wang

Verified account Nebius

@hanrui_w

SVP, Applied AI @NebiusAI | Inference & Post-train @NebiusTF | Co-Founder and CEO @Eigen_AI_Labs | PhD @MIT | Ex-@NVIDIA @XilinxInc | Efficient AI, Comp Arch
San Francisco, CA hanruiwang.com Joined December 2019
1.1K Following 4.1K Followers 468 Posts

Latest posts 5

Ryan Hanrui Wang Verified account @hanrui_w
On November 23 at 9 a.m. PT, I’ll join my colleagues to share what we’re learning about making large models faster to serve, improving them through post-training, and building agents that can handle longer tasks reliably. We’ll also talk about the problems we’re still trying to solve and where we’d like to collaborate with university researchers.

We’ll introduce the Nebius PhD Fellowship, which gives PhD students funding, GPU compute, and the chance to work directly with our research teams. If you’re working on one of these problems or advising someone who is, I’d love to hear what you’re exploring.

Register for Inside Nebius Research on November 23 at 9 a.m. PT
nebius.science/events/webinar…
Apply for the Nebius PhD Fellowship by January 15, 2027
nebius.com/phd-fellowship…
nebius.science Inside Nebius Research: Priorities, Open Questions and the PhD Fellowship Follow how AI and science co-evolve — through the lens of the cloud powering it.
Nebius Science Verified account @nebius_science 1d
What’s @nebiusai researching, and how can PhD students and professors get involved?

🎓 Join our Nov 23 webinar on our work in inference optimization, post-training, and agentic systems, and the new Nebius PhD Fellowship.

Sign up: nebius.science/events/webinar…
View quoted post on X
0 replies 2 reposts 25 likes 1.3K views View on X
Ryan Hanrui Wang Verified account @hanrui_w
When is it worth building your own inference stack?

This Wednesday, I’ll be at The Future of Inference & Compute, a private dinner in San Francisco hosted by Nebius and Timeless.

I’m looking forward to comparing notes with founders, researchers, and infrastructure engineers on what drives cost per token in production, when managed inference makes sense, and how decisions made during post-training show up later in serving.

If you’re thinking about any of these trade-offs, I’d love to hear what you’re seeing.

Wednesday, November 18 · 7–9 p.m. PT · San Francisco

luma.com/nebius-krp7
luma.com The Future of Inference & Compute · Luma The Future of Inference and Compute A private dinner hosted by Nebius and Timeless An intimate evening in San Francisco for founders, researchers, and infra…
3 replies 6 reposts 50 likes 2.3K views View on X
Ryan Hanrui Wang Verified account @hanrui_w
Inkling-2-Small, small yet powerful, now available on Nebius Token Factory from Day 0
Nebius Token Factory Verified account @nebiustf 5d
Inkling 2 Small is available on Nebius Token Factory from Day 0.

Build agents that write code, use tools, inspect visual results and iterate. @thinkymachines shows the model doing this with websites, games and 3D scenes.

Interested in running Inkling 2 Small on Nebius? Get in touch with our team to explore deployment options.
View quoted post on X
0 replies 0 reposts 24 likes 1.4K views View on X
Ryan Hanrui Wang Verified account @hanrui_w
Around six months ago, our Eigen AI team joined Nebius. Seeing what we’ve built together in that time makes this Q3 update especially meaningful to me.

In Q3, Nebius Token Factory’s contract value grew 17× quarter over quarter. And Token Factory was already approaching a multi-hundred-million-dollar run rate.

I’m especially proud to see Eigen AI's work on model optimization, quantization, inference runtimes, post-training and RL supercharged Token Factory. Together, we’re helping customers run leading open models faster and more efficiently in production. It’s been rewarding to see the work we started at Eigen AI reach more builders through Nebius.

Revolut and Higgsfield are already combining different workloads on Nebius. Applied Compute and Cursive are taking models they post-trained on Nebius AI Cloud and serving them. Hark chose Token Factory for low-latency inference in its computer-use agent, and Prosus uses it to run open models at scale across its e-commerce and lifestyle businesses.

A huge thank-you to the Eigen AI team and to the many new teammates who’ve joined us as we grow into Nebius Applied AI. I’m also grateful to everyone across Nebius who helped bring our work into Token Factory, ship new models on launch day, and support customers as they scaled.

I’m proud of what we’ve accomplished together in a few months, and I’m excited about what this growing team will build next!

@Eigen_AI_Labs @nebiusai @nebiustf
Nebius Verified account @nebiusai 5d
“Our focus remains on building a healthy, profitable business for the long term. We will continue to invest across the business — in our capacity and infrastructure build-out, in our platform and products, and in the team that takes this to market.”

– Arkady Volozh, founder and CEO

Highlights from our Q3 results:

- Customers are paying more up front. Prepayments on Q3 agreements covered more than 100% of the related capex, including data center construction and networking, up from 50–60% in Q2.
- Pricing continues to rise. Annual contract value on Q3 deals topped $25 million per MW, up from more than $20 million in Q2, and GPU-hour prices rose on average by more than 20% from Q2.
- Our customer base is growing. We added more AI-native companies and signed our first agreements with leading enterprise data and AI platforms.
View quoted post on X
12 replies 13 reposts 254 likes 22K views View on X
Ryan Hanrui Wang Verified account @hanrui_w
New blog from Nebius Applied AI. We put together a practical guide to choosing a post-training method for open-weight models.

My biggest takeaway is to start with a real task your model gets wrong, then choose a method based on the feedback you can reliably collect. If you have good examples, start with supervised fine-tuning. If people can consistently rank responses, try preference tuning. If you can score task outcomes, consider reinforcement learning. And if a stronger model can guide a smaller one, distillation is an option.

Then test the new model on tasks it hasn’t seen and measure quality, latency, and cost together. One example I liked from the blog is Plurai’s work on entity resolution. After fine-tuning Qwen3-4B on Nebius Token Factory, Plurai reported that accuracy rose from 69% to 89% while latency dropped from 1 second to 0.15 seconds per request.

A big shout-out to @kingsquare2 , Joe Sarabia and @demian_ai for leading this effort and putting the guide together

nebius.com/blog/posts/pos…
nebius.com Fine-tuning, RL, or distillation? How to choose a post-training method Choose a post-training method by its learning signal, with examples, three diagrams, and an evaluation loop for open-weight models.