Cognition Releases SWE-2, Striking a New Cost-Performance Balance for Code Models on a Kimi K3 Base

Cognition released its SWE-2 programming model on September 10, 2026, built on Moonshot AI's open-source Kimi K3 (a 2.8-trillion-parameter MoE) as a base and fine-tuned via reinforcement learning. It scores 50.0% on the FrontierCode 1.1 Main benchmark, trailing Anthropic's Fable 5.1 by less than 1 percentage point, yet runs at 64% lower cost.

Fact Recap

According to Cognition's official blog, SWE-2 is post-trained from Kimi K3, a base that has already undergone extensive agentic coding reinforcement learning. SWE-2 scores 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. Compared with Kimi K3's original performance, these represent gains of 5.8, 4.5, 4.5, and 5.8 percentage points, respectively. The model also surpasses SWE-1.7 and Grok 4.6 on multiple benchmarks and outperforms the latter on both score and cost.

Mechanism Breakdown

Training introduces a linear cost penalty mechanism that applies penalties across different effort levels within a single RL run, with penalty values adjusted according to the local slope of the base model's Pareto frontier. This approach aims to push the entire cost-performance frontier forward while preserving its shape. The reward baseline is length-weighted to stabilize training. Rollout serving boosts decoding throughput through scheduling optimization and an online draft model, combining NVFP4/FP8 kernels with quantization-aware training to achieve lower train-inference mismatch at nearly three times the parameter scale of SWE-1.7. The volume of training data was tripled, and an iterative verifier flywheel built on earlier checkpoints was established.

Industry Impact

The release demonstrates the viability of a path that pairs Chinese open-source base models with US fine-tuning teams. SWE-2 sharply reduces cost while holding scores close to the frontier, altering the cost-performance trade-off space in the code agent field. Existing models such as Fable 5.1 and GPT-5.6 Sol see their price advantage eroded at equivalent performance, and the market may reassess the commercialization threshold for trillion-parameter-scale models.

Strategic Assessment

[Analysis, not fact] RL has for the first time publicly scaled to the trillion-parameter level, which may accelerate industry validation of the scalability of large-scale reinforcement learning; if similar cost-penalty methods are widely adopted, a restructuring of the Pareto frontier for code models could reshape developers' expectations around cost-effectiveness, though real-world results still require validation across multiple scenarios.