SWE-2

Cognition

SWE-2 is Cognition's most advanced coding model, post-trained from Moonshot AI's 2.8-trillion-parameter Kimi K3 with reinforcement learning scaled to the multi-trillion-parameter regime for the first time. A single RL run trains all reasoning-effort levels (medium, high, max) together with per-effort cost penalties tuned to the base model's Pareto frontier, advancing the whole cost-performance curve rather than a single operating point. On FrontierCode 1.1 Main it scores 50.0%, within one point of Claude Fable 5.1 (50.9%) while Cognition states it is 64% cheaper; on DeepSWE 1.1 it scores 73.0%; on Terminal-Bench 2.1 it scores 92.8%; on Terminal-Bench 4.0 it scores 27.3%. SWE-2 is available starting 2026-09-10 in Devin Desktop and Devin CLI, with Devin Web and Fusion rollout in progress; Cognition has not published a standalone public API, model card, or context-window specification independent of the Devin product.

Key Specifications

Parameters
-
Context
-
Release Date
September 10, 2026
Average Score
60.8%

Timeline

Key dates in the model's history
Announcement
September 10, 2026
Last Update
October 31, 2026
Today
November 20, 2026

Technical Specifications

Parameters
-
Training Tokens
-
Knowledge Cutoff
-
Family
-
Fine-tuned from
kimi-k3
Capabilities
MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Other Tests

Specialized benchmarks
DeepSWE 1.1
DeepSWE 1.1; reasoning effort not specified in the comparison table. • Self-reported
73.0%
FrontierCode 1.1
FrontierCode 1.1 Main; weighted rubric aggregate; reasoning effort not specified in the comparison table. • Self-reported
50.0%
Terminal-Bench 2.1
Terminal-Bench 2.1; reasoning effort not specified in the comparison table. • Self-reported
92.8%
Terminal-Bench 4.0
Terminal-Bench 4.0; reasoning effort not specified in the comparison table. • Self-reported
27.3%

License & Metadata

License
proprietary
Announcement Date
September 10, 2026
Last Updated
October 31, 2026

Compare SWE-2

All comparisons

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.