
- Joined
- Dec 2025
- How they joined
- Directly
- Before Cognition
- ByteDanceStudent Researcher
- Status
- Current
“when I joined Cognition, training our own frontier-class model felt like a dream that was far away. We got there in 6 months - faster than even our most optimistic projection. With this team, truly anything feels possible.”
- Third author on 'SWE-1.7: Frontier Intelligence at a Fraction of the Cost' (2026-07-08), after co-leads Ben Pan* and Carlo Baronio*; SWE-1.7 came from RL pipeline improvements on top of Kimi K2.7, trained in Devin's harness.Jul 8, 2026
- On SWE-1.7 launch day he wrote on X: "when I joined Cognition, training our own frontier-class model felt like a dream that was far away. We got there in 6 months - faster than even our most optimistic projection. With this team, truly anything feels possible. Try SWE-1.7 with Cerebras today - you'll be blown away."Jul 8, 2026when I joined Cognition, training our own frontier-class model felt like a dream that was far away. We got there in 6 months - faster than even our most optimistic projection. With this team, truly anything feels possible. Try SWE-1.7 with Cerebras today - you'll be blown away.
- Lead (first-listed) author on Cognition's 'Introducing SWE 1.6: Improving Model UX' (2026-04-07), the model optimized for intelligence and model UX, shipped generally available in Windsurf at up to 950 tok/s with a 3-month free tier.Apr 7, 2026
- Announced on X on 2026-03-01: "I guess now is as good a time as any to announce that I finished my PhD and started as a researcher at @cognition! This has been a wild first project with an absolutely insane team, excited for what's coming next"; Cognition colleagues Ben Pan (@ybenpan) and Timothy Luong replied.Mar 1, 2026
- Quoted by Forbes in Aug 2024 as a CMU robotics PhD using Codeium's AI coding assistant, adopted 'on his former colleagues' suggestion': "Most of the time it just works out of the box." Codeium (Windsurf) was later acquired by Cognition.Aug 2024
- On FrontierCode 2.0 he wrote: "With FC2.0 we've tried to measure the models on "true" software engineering skills, and the result is a ranking that imo matches with the vibes. it turns out this isn't quite saturated yet (for now)"Oct 27, 2026With FC2.0 we've tried to measure the models on "true" software engineering skills, and the result is a ranking that imo matches with the vibes. it turns out this isn't quite saturated yet (for now)
- Co-author of Cognition's FrontierCode code-quality benchmark (2026-06) and FrontierCode 1.1 (2026-07-07) with Eric Lu, Ben Pan, Fermi Ma, Alex Lombardi, Deniz Birlikci and others; he tweeted: "Agents are writing more code than ever, yet we all viscerally know the code isn't high-quality. We worked hard to make a benchmark to quantify this "slop" aspect. We expect this benchmark will saturated sooner than later but models that perform well on it will be a joy to use."Jul 7, 2026
- On Devin Desktop he wrote: "a year ago i was hand-typing print statements and ipdb.set_trace(). today that feels quaint. devin desktop is a big step towards the future of software!"Jun 3, 2026
- Co-author of 'An Early Preview of SWE-1.6 and Research Update' (2026-03) with Carlo Baronio, Ben Pan, Sam Lee, Eric Lu, Steven Cao, Adam Zweiger and others; the preview landed around the time he publicly announced joining, matching his 'wild first project' tweet.Mar 2026
- Cognition acquired Windsurf (formerly Codeium/Exafunction), taking its IP, product, trademark and brand in July 2025, days after Google DeepMind hired Windsurf's CEO Varun Mohan; SWE-1.x models he trains ship in Windsurf (Devin Desktop).Jul 14, 2025
- Windsurf/Codeium co-founder Varun Mohan previously worked at Nuro, where Rohan was a software engineer 2019-2021.Aug 2024
- Reshared Cognition's funding announcements on LinkedIn: "proud to be a part of this!" on the $1B+ raise at a $26B valuation (2026-05-27) and "so lucky to be a part of this ride!" on the $2B+ raise at a $48B valuation (2026-09-09).Sep 9, 2026
Seed - Video Generation
Meta FAIR, Video Understanding
Behavior ML
With FC2.0 we've tried to measure the models on "true" software engineering skills, and the result is a ranking that imo matches with the vibes. it turns out this isn't quite saturated yet (for now)
when I joined Cognition, training our own frontier-class model felt like a dream that was far away. We got there in 6 months - faster than even our most optimistic projection. With this team, truly anything feels possible. Try SWE-1.7 with Cerebras today - you'll be blown away.
fusing the power of multiple models intelligently is really hard but can yield superior performance with much better cost. proud of our team for taking the lead on this! https://t.co/VGLgsVbNT9
Agents are writing more code than ever, yet we all viscerally know the code isn't high-quality. We worked hard to make a benchmark to quantify this "slop" aspect. We expect this benchmark will saturated sooner than later but models that perform well on it will be a joy to use.
try it out! I think you’ll be surprised at how fast and smart it feels.
we’re still just getting started!
That’s my friend 🥹🥹
grateful to come to this office every morning!
ppl are gonna hate but this set was actually 🔥
bro doesn’t know about pure 🥀
watch this space