Kefan XIAO

Gemini3 is out! Personally super proud of the coding capabilities in dev’s daily usage. Have been dedicating to it with @pengchengyin for the last several months and partnering with @melvinjohnsonp team to launch it!

@AndrewDai @yinfeiy @ElorianAI congrats Andrew! Best of luck and can't wait to see what you will build!

After almost 12 years in Brain/DeepMind, I’ve finally decided to take the leap. My cofounders: @yinfeiy, Seth and I have kicked-off @ElorianAI. The first multimodal reasoning lab founded and led by former LLM pretraining, data and multimodal leads. youtu.be/YlvfNpOMeOY?si… (1/n)

Sad if it’s true. Thinking machine is nothing without its people…

Pure weightlifting take: they need to add 45lbs ones. Also where can I buy some?

New Blogpost: How to game the METR plot🚨 In 2025, a single graph changed AGI timelines, investments, research priorities, model quality assessments and much more. But if you squint harder, only 14 prompts shaped AI discourse over this year. Thats all the data in the 1-4 hour horizon length regime that matters. 🕵️ What's more? A majority of these are about Cybersecurity capture the flag contests, and training a Machine Learning model. > Post-train your model on CTF and ML codebases > profit 📈! its METR horizon length will increase. Exactly what OpenAI has been targeting in its Codex model releases... and is Anthropic underperforming in the 2-4hr range because it mostly consists of cybersecurity, which is dual-use for safety? To be clear, I think its an excellent idea to track horizon lengths instead of benchmark accuracy. But under the current modelling assumption of success probability being a logistic function of task length, SWAA+HCAST accuracy improvements alone might explain the exponential progress in horizon length 🔎 In the blog, I show detailed evidence for why we need to stop overindexing on the METR plot. Share it with anyone you see making decisions based on where the latest model lands on the METR plot. shash42.substack.com/p/how-to-game-…

swebench is better than pro 🫡

Introducing Nano Banana Pro 🍌 aka Gemini 3 Pro Image, our new SOTA image generation and editing model. It is all the things you loved about @NanoBanana, but with some wild new improvements. It is available right now for developers in the Gemini API and in the Gemini App!

♊️ 🚀

@siamaksha Thank you Siamak! Great teamwork with @pengchengyin

Kudos to @KevinKiao, he did lots of the heavy lifting for swebench and more.

New SOTA on official swebench leaderboard with Gemini 3 pro! swebench.com We carefully designed our RL so it works without overfitting. Please try the model out and give us feedbacks!

I’m especially proud of where we landed on coding and agentic use cases. Looking at the charts for Terminal-bench 2.0, SWE-Bench and 2-Bench compared to Gemini 2.5 shows the incredible jump, but using it to actually solve hard problems is the real win.

@melvinjohnsonp congrats Melvin! Has been a great push!

Been waiting a long time to share this one. Meet Gemini 3 Pro. It’s our most intelligent multimodal model that’s deeply capable. x.com/sundarpichai/s…

@sunjiao123sun_ Great work!

In the past two months, our small Webapp Coding team have been cooking hard to make Gemini great at WedDev, and we are thrilled to claim the 👑! Yes, we saw your enthusiasm — pelican riding a bike, game controller, please keep trying and sending your best WebDev prompts to our the way! We love them! Besides Webdev Arena, we also achieved #1 on Design Arena across categories: website gen, game gen, ui component gen etc! Website lovers, designers, we can’t wait to hear your feedback!

And it’s great at software engineering too!

@Skiminok @pengchengyin @melvinjohnsonp Thanks Alex! You had been pushing hard on this as well!

@KevinKiao @pengchengyin @melvinjohnsonp Such an impressive model, congrats!

Gemini3 is out! Personally super proud of the coding capabilities in dev’s daily usage. Have been dedicating to it with @pengchengyin for the last several months and partnering with @melvinjohnsonp team to launch it!

@_arohan_ @pengchengyin @melvinjohnsonp Thanks Rohan!

@KevinKiao @pengchengyin @melvinjohnsonp Congrats!

@ankesh_anand Has been a great journey!🫡💪

Gemini3 Pro is out, very exciting to be able to push the frontier with this one! There was never a dull day post-training this model, I hope the combination of a strong base model with sota reasoning is evident! This is obviously a big leap compared to 2.5 Pro, but I am excited about our research agenda more than ever. The models will continue to get smarter!

And this has been a great team work with many friends!

One behavior I really like of gemini3 pro is that it actively uses tools to explore and verify. And it has been helping my daily works! Please apply it in your workflow and tell us how do you feel!

@_mohansolo Congrats Varun!

Excited to launch Google Antigravity, our next generation agentic IDE, now powered by Gemini 3!