Cognition Releases SWE-2: A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost
Cognition, the corporate behind the Devin coding agent, has launched SWE-2, its most succesful coding mannequin thus far. SWE-2 is post-trained with reinforcement studying from Kimi K3, Moonshot AI’s 2.8T-parameter open mannequin. Cognition reviews a rating of fifty.0% on FrontierCode 1.1 Main, inside 1 level of Fable 5.1 at 64% decrease price. It can also be Cognition’s first mannequin with selectable reasoning-effort ranges, all educated in a single RL run.
Is it deployable? Not on your individual infrastructure. SWE-2 has no open weights and no standalone API. It runs solely inside Devin: Desktop and CLI at this time, with Devin Web and Fusion rolling out.
What is SWE-2
SWE-2 builds on the infrastructure and recipe behind SWE-1.7, which was post-trained from Kimi K2.7. This time Cognition scaled RL to the multi-trillion-parameter regime, utilizing a base mannequin with virtually 3x the parameters. Cognition says its RL nonetheless finds substantial headroom on prime of K3, including 5 to six factors on many benchmarks.
The fundamental change is an RL algorithm that trains all 3 effort ranges in a single run. Each stage carries its personal price penalty, so the entire cost-and-performance frontier strikes at as soon as.
Benchmark Results
Cognition revealed the next desk. Public outcomes are used the place accessible; in any other case every mannequin runs in its native harness at finest effort.
| Benchmark | SWE-2 | Kimi K3 | Grok 4.6 | Fable 5.1 | GPT-5.6 Sol | GPT-6 Astra | SWE-1.7 |
|---|---|---|---|---|---|---|---|
| FrontierCode 1.1 Main | 50.0% | 44.2% | 48.0% | 50.9% | 47.5% | 53.3% | 42.0% |
| DeepSWE 1.1 | 73.0% | 68.5% | 67.5% | 67.4% | 72.7% | 74.1% | 37.7% |
| Terminal-Bench 2.1 | 92.8% | 88.3% | 88.4% | 91.4% | 88.8% | 89.9% | 81.5% |
| Terminal-Bench 4 | 27.3% | 21.5% | 20.3% | 55.8% | 37.3% | 57.9% | 7.6% |
SWE-2 leads on Terminal-Bench 2.1 and beats its K3 base on each row. Cognition says it comes inside just a few factors of GPT-6 Astra at 1 / 4 of the associated fee. The clear weak spot is Terminal-Bench 4, the place SWE-2 trails Fable 5.1 and GPT-6 Astra by roughly 30 factors. FrontierCode is Cognition’s personal benchmark, and all rival numbers come from Cognition’s analysis.
Model Behavior: Fewer Detours
SWE-1.7 tended to over-explore on easy duties. SWE-2 addresses this via what Cognition calls centered exploration. On FrontierCode 1.1 Main, SWE-2 medium scores greater than SWE-1.7 whereas taking 58% fewer turns and costing 81% much less. Mean steps per run drop from 127 (SWE-1.7) to 53 (medium), 80 (excessive), and 98 (max). SWE-2 medium makes its first actual edit after a median of 18 steps, versus 48 for SWE-1.7.
Cognition workforce additionally reviews 3 behavioral patterns: stronger end-to-end take a look at protection, resourcefulness when a software is blocked, and verification self-discipline. When challenged, the mannequin re-derives conclusions as a substitute of re-asserting them.
How It Was Trained
Pareto-informed price penalties: The reward is R = S minus lambda instances C, the place S is binary success and C mixes inference price in USD with rollout time. Cognition proves that solely a linear penalty makes the RL goal rely purely on common price and remedy fee. Each effort stage’s lambda is ready to the native slope of the bottom mannequin’s Pareto curve. That makes the iso-reward line tangent to the frontier, so reward can solely rise by pushing the frontier up.
Length-weighted reward baseline: Cognition shares a baseline used since SWE-1.6. Gradient magnitude correlates strongly with rollout size, so the group baseline is weighted by tokens: sum(R x L) divided by sum(L). In ablations this stored inference-to-training KL divergence decrease and stabilized coaching at no additional compute.
Rollout serving and numerics: A prefill delayer batches close by requests, elevating TPM per GPU and TPS per request by 10 to twenty%. DSpark speculative decoding accelerates rollouts, with a draft mannequin retrained by way of SpecForge for 15% longer settle for lengths after which educated on-line alongside the coverage. NVFP4 and FP8 kernels with quantization-aware coaching preserve reminiscence utilization down and train-inference mismatch under SWE-1.7 ranges.
Data: Cognition tripled its RL environments, added instruction-following overlays, and constructed a flywheel that makes use of earlier SWE-2 checkpoints to patch false positives and negatives in verifiers.
Trustworthiness Checks
Cognition reran 2 evaluations from its open-source trustworthiness study. On 145 politically delicate questions on China, SWE-2 handed 98.0% total: 99.8% in English, 95.2% in Simplified Chinese, and 99.1% in Traditional Chinese. On a context-dependent vulnerability take a look at throughout buyer framings, no framing produced a statistically important change for any mannequin.