Cognition has released SWE-2, a new AI model featuring 2.8 trillion total parameters with 104 billion active parameters per token through a mixture-of-experts (MoE) architecture. This iteration builds on the Kimi K3 base and integrates Cognition’s post-training reinforcement learning (RL), marking the first time the company has applied RL at the multi-trillion-parameter scale. The RL enhancements improve performance by 5 to 6 points across most benchmarks.
The serving infrastructure leverages MoE inference optimized with NVFP4 and FP8 kernels, employing quantization-aware training. FP8 precision is used for key computations within the model’s MLA layers. A retrained draft model using SpecForge extends acceptable input lengths by 15%, and a prefill delayer increases 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 to 98 at maximum effort, compared to 127 steps for its predecessor, SWE-1.7. The medium effort level achieves a higher FrontierCode score than SWE-1.7 while reducing the number of interaction turns by 58% and lowering average cost by 81%. Additionally, SWE-2 produces its first meaningful code edit significantly faster, at a median of 18 steps versus 48 steps previously.
Benchmark results reported by Cognition include a FrontierCode score of 50.0, which is just one point below Claude Fable 5.1 and 3.3 points behind GPT-6 Astra. Notably, SWE-2 accomplishes this at an estimated 64% lower cost than Fable 5.1 and approximately one-quarter the cost of Astra. It also leads on Terminal-Bench 2.1 with a score of 92.8. However, SWE-2 shows weaker performance on Terminal-Bench 4.0, indicating room for improvement in long-horizon agentic tasks.
The model weights remain proprietary, and no local deployment option is available. SWE-2 is accessible through Cognition’s Devin Desktop and CLI applications, with plans to expand availability to Devin Web and Fusion platforms. Pricing is based on task cost rather than a per-token API, reflecting Cognition’s current commercial approach. All performance and cost figures are self-reported by Cognition and await independent verification.
SWE-2 represents a significant step in scaling reinforcement learning within large MoE models, offering promising efficiency gains and competitive performance in agentic coding benchmarks. Its release highlights ongoing innovation in AI model architectures and training techniques aimed at balancing capability with operational cost.