{"version":1,"type":"story","url":"https://digestai.news/story/cognition-s-swe-2-breaks-terminal-bench-2-1-with-92-8","json":"https://digestai.news/story/cognition-s-swe-2-breaks-terminal-bench-2-1-with-92-8.json","markdown":"https://digestai.news/story/cognition-s-swe-2-breaks-terminal-bench-2-1-with-92-8.md","slug":"cognition-s-swe-2-breaks-terminal-bench-2-1-with-92-8","headline":"Cognition Releases SWE-2 Coding Model Trained on Kimi K3","summary":"Cognition has launched SWE-2, its most advanced coding model to date, post-trained using reinforcement learning on Moonshot AI's 2.8-trillion-parameter open model, Kimi K3. The model achieves a 50.0% score on the FrontierCode 1.1 Main benchmark, placing it within one percentage point of Fable 5.1 while operating at a 64% lower cost. \n\nSWE-2 introduces selectable reasoning-effort levels (medium, high, and max) trained within a single reinforcement learning run using Pareto-informed cost penalties. Compared to its predecessor, SWE-1.7, the new model significantly reduces over-exploration, with the medium setting taking 58% fewer turns and costing 81% less on benchmarks. SWE-2 is not available as open weights or a standalone API; it is accessible exclusively through Cognition's Devin platform.","keyPoints":["SWE-2 scores 50.0% on FrontierCode 1.1 Main, matching Fable 5.1 performance at a 64% lower cost.","The model is post-trained via reinforcement learning from Moonshot AI's 2.8T-parameter Kimi K3 base.","It is available exclusively within the Devin agent platform, with no standalone API or open weights."],"whyItMatters":"SWE-2 demonstrates that scaling reinforcement learning on top of massive open-weights models can yield frontier-class coding capabilities at a fraction of the cost of proprietary giants, shifting the economics of AI-assisted software engineering.","category":{"slug":"models","name":"Generative AI & Models","url":"https://digestai.news/category/models"},"entities":{"companies":["Cognition","Moonshot AI"],"models":["SWE-2","Kimi K3","Fable 5.1","SWE-1.7","GPT-6 Astra"],"people":[]},"firstPublishedAt":"2026-09-10T16:52:11Z","updatedAt":"2026-09-12T23:56:03Z","sourceCount":2,"hasPrimarySource":false,"sources":[{"outlet":"MarkTechPost","title":"Cognition Releases SWE-2: A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost","url":"https://marktechpost.com/2026/09/12/cognition-releases-swe-2-a-kimi-k3-post-trained-coding-model-that-matches-fable-5-1-on-frontiercode-at-64-lower-cost","publishedAt":"2026-09-12T23:56:03Z","type":"press","primary":false,"lead":true},{"outlet":"tokenstead.ai","title":"Cognition's SWE-2 achieves 92.8 on Terminal-Bench 2.1","url":"https://tokenstead.ai/models/swe-2","publishedAt":"2026-09-10T16:52:11Z","type":"press","primary":false,"lead":false}],"sourceNotes":null,"discussions":[{"site":"Hacker News","url":"https://news.ycombinator.com/item?id=49646778","points":61}],"thread":null,"cite":{"text":"Digest AI, \"Cognition Releases SWE-2 Coding Model Trained on Kimi K3\", 10 September 2026, https://digestai.news/story/cognition-s-swe-2-breaks-terminal-bench-2-1-with-92-8","publisher":"Digest AI","title":"Cognition Releases SWE-2 Coding Model Trained on Kimi K3","datePublished":"2026-09-10T16:52:11Z","url":"https://digestai.news/story/cognition-s-swe-2-breaks-terminal-bench-2-1-with-92-8"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}