Introducing SWE-2: Pushing the Pareto Frontier

gist

Cognition introduces SWE-2, a Devin coding model post-trained from Kimi K3 (2.8T) that reaches 50.0% on FrontierCode 1.1 Main—within a point of Fable 5.1 at 64% lower cost. One RL run trains every effort level with Pareto-slope cost penalties, a length-weighted reward baseline, and hardened verifiers, so efficiency and capability move together: SWE-2 medium edits sooner, explores less, and costs far less than SWE-1.7.

ideas

• One RL run advances the whole cost–performance frontier. Linear cost penalties per effort level are set to the base model’s local Pareto slope so reward gains lift the frontier instead of collapsing high effort into medium.
• Focused exploration beats thorough overthinking. SWE-2 medium scores higher than SWE-1.7 on FrontierCode while using 58% fewer turns and 81% less cost, with first real edits at median 18 steps versus 48.
• Length-weighted baselines stabilize multi-trillion RL. Weighting the group baseline by rollout length approximates the optimal variance-reducing baseline without an extra backward pass and keeps inference–training KL low.
• Verifier flywheels matter as models get resourceful. Tripled RL environments plus instruction overlays and iterative patching of false positives and negatives from prior SWE-2 checkpoints curb reward hacking on a stronger Kimi K3 base.
• Trust and framing checks ride with the release. Updated propaganda/censorship and context-dependent vulnerability evals report high pass rates and no significant framing effect on vulnerability for SWE-2 or peers.

quotes