Cognition releases SWE-2, its in-house coding model built on Kimi K3

Archive entry: this covers an event from September 10, 2026, and we published it on November 8, 2026. The details come from sources dated at the time and may have changed since.

On September 10, 2026, Cognition released SWE-2, its most advanced in-house coding model, and described it in its announcement post, credited to The Cognition Team. SWE-2 was post-trained from Moonshot's Kimi K3, a model with 2.8 trillion parameters.

What changed

• Model: SWE-2 is post-trained from Kimi K3 (2.8T parameters). Cognition said it was the first time it had taken reinforcement learning (RL) to multi-trillion-parameter scale, with a single RL run covering all reasoning-effort levels.
• Scores: Cognition reported 50.0% on FrontierCode 1.1 Main, 73.0% on DeepSWE 1.1, 92.8% on Terminal-Bench 2.1 and 27.3% on Terminal-Bench 4.0. All of these are self-reported.
• Versus Fable 5.1: on FrontierCode 1.1 Main, Cognition said SWE-2 was within one point of Fable 5.1 (which scored 50.9%) while being 64% cheaper.
• Versus SWE-1.7: on the same benchmark, SWE-2 medium took about 58% fewer steps and had an 81% lower average cost.

What it means for people who build with AI

For people working in Devin Desktop and Devin CLI, SWE-2 was a cheaper in-house model that Cognition said came close to frontier models on coding benchmarks. Because every score in the post was reported by Cognition itself, the figures describe the company's own claims rather than independent testing.

Availability

SWE-2 was available on the day of the announcement in Devin Desktop and Devin CLI, and Cognition said it was rolling out on Devin Web and Fusion. The post gave no per-token price, context window or plan details.

Sources

• Introducing SWE-2: Pushing the Pareto Frontier (Cognition, September 10, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.