Research · Researcher
Ryan Kim
- Joined
- Not public
- How they joined
- Directly
- Status
- Current
Details
Company sitecognition.ai
Notes
- Listed as a co-author (fifth in the byline: Ben Pan*, Carlo Baronio*, Rohan Choudhury, Eric Lu, Ryan Kim, Deniz Birlikci, TC Qin, Sam Lee, Fermi Ma, ...) of Cognition's SWE-1.7 launch post, the RL-trained coding model built on a Kimi K2.7 base.Jul 8, 2026
- Cognition followed SWE-1.7 with SWE-2 (post-trained on Kimi-K3) in September 2026, saying its RL recipe keeps scaling on stronger base models. The sources reviewed do not confirm whether Ryan Kim is credited on SWE-2, but it is the next model from the RL post-training program he worked on.Sep 10, 2026
- SWE-1.7 was trained from a Kimi K2.7 base that had already undergone extensive RL post-training.Jul 8, 2026
- SWE-1.7 scores 42.3% on FrontierCode 1.1 MainScore.Jul 8, 2026
- SWE-1.7 scores 81.5% on Terminal-Bench 2.1.Jul 8, 2026
- SWE-1.7 scores 77.8% on SWE-Bench Multilingual.Jul 8, 2026
- On FrontierCode 1.1 MainScore, SWE-1.7's 42.3% trails GPT-5.5 at 43.0% and Opus 4.8 at 46.5% but leads Kimi K2.7 at 30.1%, Opus 4.7 at 38.5%, Composer 2.5 at 25.6%, GLM 5.2 at 24.5%, and SWE-1.6 at 9.4%.Jul 8, 2026
- Cognition claims SWE-1.7 reaches frontier-level intelligence at a much lower cost, advancing the cost-performance Pareto curve.Jul 8, 2026
- Training techniques included top-p sampling with sampling distribution replay to stave off entropy collapse and bound training-inference divergence.Jul 8, 2026
- Cognition used the Muon optimizer and eliminated non-deterministic operations in the trainer.Jul 8, 2026
- RL training spanned four datacenters across three continents, combining Cognition's own GPUs with additional compute from inference providers like Fireworks.Jul 8, 2026
- Compressed weight deltas between the current and previous weights reduced the size of each transfer by over 99%.Jul 8, 2026
Articles
