Cognition Releases SWE-2 Flash, a 276B Coding Model Outperforming Larger American Open-Weights

• Cognition introduces SWE-2 Flash, a 276B total / 12B active parameter coding model that outperforms all American open-weight models, including those 3x larger, on leading evaluations.
• The model improves Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0% over the base Inkling-Small model, while raising ExploitBench scores from 11.3% to 34.8%.
• SWE-2 Flash is available in Devin Desktop and CLI at one-fifth the price and 3x the speed (200 tok/s vs 60 tok/s) of the larger SWE-2 model.
• Weights are not publicly released due to cyber risks; Cognition is building a program to securely share them with trusted industry partners and research labs.

Training Methodology

• A new mid-training stage was added before reinforcement learning, using a few trillion tokens of high-quality scientific articles, blogs, and code.
• Data mixtures were optimized using token efficiency plots to maximize validation loss improvement per token.
• Reinforcement learning focused on high-quality, diverse tasks to build general reasoning and problem-solving skills for real-world engineering.

Availability and Access

• Free access is available through November 30 in Devin Desktop and CLI.
• The model is designed for on-premises deployment to meet enterprise and government needs for observability and control.