SWE-2 Flash

Cognition

SWE-2 Flash is Cognition's fast, low-cost coding model, post-trained on Thinking Machines' Inkling-Small (276B total / 12B active) with a multi-trillion-token mid-training stage and large-scale reinforcement learning on coding and agentic tasks. Cognition reports 60.1% on DeepSWE 1.1, 45.5% on FrontierCode 1.1 Main, 24.0% on FrontierCode 2.0, 89.2% on Terminal-Bench 2.1 and 20.0% on Terminal-Bench 4.0. It runs in Devin Desktop and Devin CLI (free through 2026-11-30) at one-fifth the price and roughly 3x the generation speed of SWE-2; weights are not public and no standalone API or per-token pricing has been published.

Key Specifications

Parameters
276.0B
Context
-
Release Date
November 4, 2026
Average Score
46.2%

Timeline

Key dates in the model's history
Announcement
November 4, 2026
Last Update
November 12, 2026
Today
November 20, 2026

Technical Specifications

Parameters
276.0B
Training Tokens
-
Knowledge Cutoff
-
Family
-
Fine-tuned from
inkling-small
Capabilities
MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Other Tests

Specialized benchmarks
DeepSWE 1.1
DeepSWE 1.1; from Cognition's SWE-2 Flash comparison table; reasoning effort not specified. • Self-reported
60.1%
ExploitBench
ExploitBench; from Cognition's SWE-2 Flash comparison table; reasoning effort not specified. • Self-reported
34.8%
FrontierCode 1.1
FrontierCode 1.1 Main; from Cognition's SWE-2 Flash comparison table; reasoning effort not specified. • Self-reported
45.5%
FrontierCode 2.0
FrontierCode 2.0; from Cognition's SWE-2 Flash comparison table; reasoning effort not specified. • Self-reported
24.0%
SWE Atlas - Codebase QnA
SWE-Atlas QA; from Cognition's SWE-2 Flash comparison table; reasoning effort not specified. • Self-reported
50.0%
Terminal-Bench 2.1
Terminal-Bench 2.1; from Cognition's SWE-2 Flash comparison table; reasoning effort not specified. • Self-reported
89.2%
Terminal-Bench 4.0
Terminal-Bench 4.0; from Cognition's SWE-2 Flash comparison table; reasoning effort not specified. • Self-reported
20.0%

License & Metadata

License
proprietary
Announcement Date
November 4, 2026
Last Updated
November 12, 2026

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.