Agents

Cognition Releases SWE-2 Coding Model Based on Kimi K3

Cognition has released SWE-2, post-training Kimi K3 to match leading coding benchmarks while lowering inference costs across multiple effort levels.

A
AIDeveloper44 Team
September 10, 2026·4 min read
Cognition Releases SWE-2 Coding Model Based on Kimi K3

SWE-2 leverages multi-trillion parameter post-training to optimize code generation efficiency.

TL;DR
  • Cognition has launched SWE-2, a coding model post-trained from the 2.8-trillion-parameter Kimi K3 base model.
  • The model introduces a single-run reinforcement learning technique that trains multiple reasoning effort levels while penalizing inference costs.
  • On FrontierCode 1.1 Main, SWE-2 achieved a 50.0% score, approaching frontier models while taking significantly fewer exploration steps.
  • SWE-2 is deployed in Devin Desktop and CLI, with rollouts underway for Devin Web and Fusion.

Overview of SWE-2

AI software company Cognition announced the release of SWE-2, its latest agentic coding model designed to improve performance on software engineering benchmarks while reducing inference expenditure. SWE-2 is post-trained from Kimi K3, an open-weights 2.8-trillion-parameter base model that previously underwent agentic coding reinforcement learning (RL).

According to Cognition's benchmark reports, the post-training process yields an additional 5 to 6 points across standardized software evaluations, optimizing the cost-performance Pareto frontier across distinct reasoning effort levels.

Benchmark Results and Efficiency Gains

On the FrontierCode 1.1 Main evaluation, SWE-2 reached a 50.0% success rate. This places the model within 0.9 percentage points of Fable 5.1 (50.9%), while Cognition notes it operates at a 64% lower cost. Compared to earlier models, SWE-2 outperformed SWE-1.7 (42.0%), Kimi K3 (44.2%), and Grok 4.6 (48.0%), and scored higher than GPT-5.6 Sol (47.5%). It finished a few points behind GPT-6 Astra (53.3%), while running at approximately one-quarter of the cost.

Additional benchmark evaluations shared by Cognition show performance across other engineering suites:

  • DeepSWE 1.1: SWE-2 scored 73.0%, compared to 68.5% for Kimi K3, 72.7% for GPT-5.6 Sol, and 74.1% for GPT-6 Astra.
  • Terminal-Bench 2.1: SWE-2 registered 92.8%, edging out Fable 5.1 (91.4%) and GPT-6 Astra (89.9%).
  • Terminal-Bench 4: SWE-2 recorded 27.3%, lagging behind Fable 5.1 (55.8%) and GPT-6 Astra (57.9%), but improving over Kimi K3's 21.5% and SWE-1.7's 7.6%.

Behavioral Adjustments and Focused Exploration

Cognition reported that user feedback regarding SWE-1.7 highlighted excessive exploration and overthinking on routine development tasks. In response, SWE-2 aims for more focused codebase analysis before executing edits. On FrontierCode 1.1 Main, the medium-effort configuration of SWE-2 averaged 53 total steps per run, compared to 127 steps for SWE-1.7—representing a 58% reduction in interaction turns and an 81% reduction in average cost.

SWE-2 medium executes its first file modification after a median of 18 steps, whereas SWE-1.7 required a median of 48 steps. The company stated that higher model capacity allows the agent to filter irrelevant directory branches earlier. Cognition also observed improvements in test generation discipline, where the model writes end-to-end integration tests to detect regressions and re-derives technical conclusions rather than repeating incorrect assumptions.

Post-Training and Optimization Recipe

To train multiple reasoning configurations without using multiple separate models, Cognition implemented a single RL training pipeline covering all reasoning effort levels. Key technical details highlighted in the announcement include:

  • Cost-Penalized Reward Functions: Cognition applied an objective formulation R = S - λeC, where S represents task success, C measures combined runtime and API cost, and λe scales with the local tangent of the base model's Pareto frontier at effort level e.
  • Reward Baselines: Training incorporated length-weighted reward baselines to stabilize gradient variance across varying trajectory lengths.
  • Serving Infrastructure: To handle the 2.8T parameter base model efficiently, Cognition deployed an online draft model to raise decoding throughput, alongside NVFP4 and FP8 quantization-aware kernels.
  • Expanded Data Flywheel: The training environment count was tripled, incorporating instruction-following overlays and automated verifiers hardened against previous model checkpoints.

Availability and Integration

SWE-2 has been made available immediately in Cognition's Devin Desktop application and command-line interface (CLI). The company confirmed that rollouts to Devin Web and the Devin Fusion collaborative workspace are currently ongoing.

Diagram: Architecture and deployment flow of Cognition's SWE-2 coding model, post-trained from Kimi K3 with configurable inference effort levels.

Enjoyed this?

Get more posts like this delivered to your inbox.