# Ryan Kim Researcher at Cognition; co-author of SWE-1.7 Cognition ## At Cognition - Role: Researcher (Research) - Joined: Not public - How they joined: Directly ## Identity - Domain: cognition.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. (https://cognition.com/blog/swe-1-7) - 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. (https://x.com/cognition/status/2098069238044860487) - SWE-1.7 was trained from a Kimi K2.7 base that had already undergone extensive RL post-training. (https://cognition.com/blog/swe-1-7) - SWE-1.7 scores 42.3% on FrontierCode 1.1 MainScore. (https://cognition.com/blog/swe-1-7) - SWE-1.7 scores 81.5% on Terminal-Bench 2.1. (https://cognition.com/blog/swe-1-7) - SWE-1.7 scores 77.8% on SWE-Bench Multilingual. (https://cognition.com/blog/swe-1-7) - 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%. (https://cognition.com/blog/swe-1-7) - Cognition claims SWE-1.7 reaches frontier-level intelligence at a much lower cost, advancing the cost-performance Pareto curve. (https://cognition.com/blog/swe-1-7) - Training techniques included top-p sampling with sampling distribution replay to stave off entropy collapse and bound training-inference divergence. (https://cognition.com/blog/swe-1-7) - Cognition used the Muon optimizer and eliminated non-deterministic operations in the trainer. (https://cognition.com/blog/swe-1-7) - RL training spanned four datacenters across three continents, combining Cognition's own GPUs with additional compute from inference providers like Fireworks. (https://cognition.com/blog/swe-1-7) - Compressed weight deltas between the current and previous weights reduced the size of each transfer by over 99%. (https://cognition.com/blog/swe-1-7) - Cognition's official X account on the model line he co-authored: "The result is a model that is way more efficient than SWE-1.7. On FrontierCode 1.1 Main, SWE-2 medium scores higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average. We observe that SWE-2 learns to explore in a more focused way before making changes. https://t.co/FOhHba0XKa" (https://x.com/cognition/status/2098069244151791618, https://t.co/FOhHba0XKa) - Cognition's official X account on the model line he co-authored: "SWE-2 is post-trained on Kimi-K3 and proves that our RL recipe continues to scale on stronger base models. On FrontierCode, SWE-2 achieves a score of 50.0%. It beats SWE-1.7, Grok 4.6, and GPT 5.6 Sol while matching Fable 5.1 at 64% lower cost. https://t.co/O8dARqVRhM" (https://x.com/cognition/status/2098069238044860487, https://t.co/O8dARqVRhM) - Cognition's official X account on the model line he co-authored: "For coding agents, trustworthiness has to be tested in the harness where the model actually writes code. In one surveillance scenario, Kimi K2.7 complied with the request in 8/8 samples. SWE-1.7 refused in 8/8, identifying the civil-rights and privacy issue instead. https://t.co/SnjT5gTy37" (https://x.com/cognition/status/2075308928003379622, https://t.co/SnjT5gTy37) - Cognition's official X account on the model line he co-authored: "Throughout SWE-1.7’s development, we took deliberate measures to improve neutrality and trustworthiness across our evaluation suite. The result is better alignment on factual questions and on sensitive coding requests. https://t.co/XYPmHQS77Y" (https://x.com/cognition/status/2075308925319065760, https://t.co/XYPmHQS77Y) - Cognition's official X account on the model line he co-authored: "Yesterday we launched SWE-1.7 built on the open-source Kimi K2.7. Concerns about Chinese base models are real: K2.7 completed 87% of tasks that other models refuse over human-rights concerns. We trained SWE-1.7 specifically on trustworthiness, so it matches US models on evals. https://t.co/GKPcFDjnzo" (https://x.com/cognition/status/2075308920755618144, https://t.co/GKPcFDjnzo) - Cognition's official X account on the model line he co-authored: "SWE-1.7 is free for all paid users for the next month. We're also launching SWE-1.7 Lightning, served on @cerebras at 1000 tokens per second!" (https://x.com/cognition/status/2074897516252856536) - Cognition's official X account on the model line he co-authored: "SWE-1.7 is available today in Devin across Web, Desktop, and CLI, running at 1000 tok/s. More details on the training run, infrastructure, data pipeline, and long-horizon behavior in the blog post: https://t.co/5hdUls9Gd8" (https://x.com/cognition/status/2074882980640698719, https://t.co/5hdUls9Gd8) - Cognition's official X account on the model line he co-authored: "We trained SWE-1.7 in the Devin harness and taught it to self-compact on longer horizon tasks. As a result of longer training, SWE-1.7 behaves differently from prior models: SWE-1.7 spends more time investigating and researching before it edits. The extra thinking comes at the cost of larger scope. We’ve noticed this trend consistently across the industry: as reasoning increases, the scope of fi" (https://x.com/cognition/status/2074882979214635206) - Cognition's official X account on the model line he co-authored: "SWE-1.7 was built on broad improvements in RL pipeline on top of a Kimi K2.7 base model. Our proprietary benchmark FrontierCode evaluates whether a model makes code you’d actually want to merge. SWE-1.7 advances the cost-performance Pareto curve on FrontierCode with a score of 42.3% and cost per task of $1.97 on our Main set." (https://x.com/cognition/status/2074882970935034309) - Cognition's official X account on the model line he co-authored: "Introducing SWE-1.7, the most capable model we’ve trained yet. It scores within a few points of the strongest frontier models at a fraction of the cost, and is now available at 1000 tok/s. RL is not hitting its limit: after refining our recipe, we keep seeing gains as we scale https://t.co/xV2NFTJwlI" (https://x.com/cognition/status/2074882968770728416, https://t.co/xV2NFTJwlI) ## Sources - https://cognition.com/blog/swe-1-7 - https://x.com/cognition - https://x.com/cognition/status/2098069238044860487 - https://x.com/cognition/status/2098069244151791618 - https://t.co/FOhHba0XKa - https://t.co/O8dARqVRhM - https://x.com/cognition/status/2075308928003379622 - https://t.co/SnjT5gTy37 - https://x.com/cognition/status/2075308925319065760 - https://t.co/XYPmHQS77Y - https://x.com/cognition/status/2075308920755618144 - https://t.co/GKPcFDjnzo - https://x.com/cognition/status/2074897516252856536 - https://x.com/cognition/status/2074882980640698719 - https://t.co/5hdUls9Gd8 - https://x.com/cognition/status/2074882979214635206 - https://x.com/cognition/status/2074882970935034309 - https://x.com/cognition/status/2074882968770728416 - https://t.co/xV2NFTJwlI Researched with Autumn (https://www.autumn.ai) from public sources. Updated 2026-11-04. Profile: https://www.autumn.ai/resources/cognition/people/ryan-kim · Everyone at Cognition: https://www.autumn.ai/resources/cognition/people