Cognition has launched SWE-2, a new AI model featuring 2.8 trillion total parameters and 104 billion active parameters per token through a mixture-of-experts (MoE) architecture. This model builds on the Kimi K3 base and integrates Cognition’s post-training reinforcement learning, marking the company’s first application of RL at the multi-trillion-parameter scale. The reinforcement learning enhancements have added 5 to 6 points on most performance benchmarks, particularly improving agentic coding tasks.
The serving infrastructure for SWE-2 employs MoE inference optimized with NVFP4 and FP8 kernels, alongside quantization-aware training. FP8 precision is used for key computations within the model’s MLA layers, including key, query, value, and scoring operations. A draft retrained version using SpecForge extends acceptable input lengths by 15%, while a prefill delayer improves throughput per GPU and tokens processed per second by 10 to 20%, albeit with some trade-offs in time-to-first-token latency.
SWE-2 supports multiple 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. The medium effort setting achieves a higher FrontierCode score than SWE-1.7 while requiring 58% fewer interaction turns and reducing average cost by 81%. Additionally, the model typically produces its first meaningful edit by step 18, significantly earlier than SWE-1.7’s step 48.
Benchmark results reported by Cognition show SWE-2 scoring 50.0 on FrontierCode 1.1 Main, narrowly behind Claude Fable 5.1 (50.9) and GPT-6 Astra (53.3). Notably, SWE-2 achieves this at an estimated 64% lower cost than Fable 5.1 and about one-quarter the cost of Astra. On Terminal-Bench 2.1, SWE-2 leads with a score of 92.8, the highest in the published comparisons. However, the model underperforms on Terminal-Bench 4.0, with a score of 27.3, trailing the top models by a significant margin, indicating room for improvement in long-horizon agentic tasks.
SWE-2’s weights remain proprietary and are not available for local deployment or download. The model is accessible through Cognition’s Devin Desktop and CLI platforms, with planned rollouts on Devin Web and Fusion. Cognition does not provide a per-token API pricing model for SWE-2; instead, cost comparisons are based on task pricing, reflecting the company’s internal metrics pending independent verification.
The introduction of SWE-2 highlights ongoing advancements in scaling reinforcement learning within large-scale MoE architectures, offering a balance of performance and cost efficiency that could influence future AI development and deployment strategies.