Cognition has released SWE-2, a large-scale mixture-of-experts (MoE) language model with 2.8 trillion total parameters and 104 billion active parameters per token. This model builds on the Kimi K3 base and incorporates Cognition’s post-training reinforcement learning (RL) enhancements, marking the company’s first application of RL at the multi-trillion-parameter scale. The RL training further improved performance by 5 to 6 points across several benchmarks.

SWE-2’s serving infrastructure uses MoE inference optimized with NVFP4 and FP8 kernels, supporting quantization-aware training. The FP8 precision handles key computations in the model’s MLA layers, including key, query, value, and scoring operations. A retrained draft model with SpecForge extends acceptable input lengths by 15%, while a prefill delayer boosts throughput per GPU and tokens per second by 10 to 20%, albeit with some tradeoffs in time-to-first-token.

The model operates at varying effort levels, with mean steps per run ranging from 53 at medium effort to 98 at maximum effort, compared to 127 steps for its predecessor SWE-1.7. Notably, the medium effort setting achieves a higher FrontierCode score than SWE-1.7 while using 58% fewer turns and reducing average cost by 81%. It also reaches its first meaningful edit much earlier in the process.

Benchmark results reported by Cognition show SWE-2 scoring 50.0 on FrontierCode 1.1 Main, just behind Claude Fable 5.1 and GPT-6 Astra, but at a claimed 64% lower cost than Fable 5.1 and one-quarter the cost of Astra. SWE-2 leads on Terminal-Bench 2.1 with a score of 92.8 but trails on Terminal-Bench 4.0, indicating room for improvement in long-horizon agentic tasks.

The SWE-2 model weights remain proprietary and are not available for local deployment. Access is currently provided through Cognition’s Devin Desktop and CLI platforms, with plans to extend availability to Devin Web and Fusion. Pricing and cost comparisons are based on Cognition’s internal metrics rather than a public API rate card.

This release highlights Cognition’s advancements in scaling reinforcement learning techniques within massive MoE models, offering a balance of competitive performance and cost efficiency. It underscores ongoing efforts to push the boundaries of agentic AI capabilities in complex coding and reasoning tasks.