Cognition launches SWE-2 Flash, a fast, low-cost coding model

What Cognition is claiming

According to Cognition, SWE-2 Flash "sets a new bar for intelligence at low cost and high speed." The company says it added a mid-training stage before reinforcement learning for the first time. Cognition reports that this improved Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%, compared with the Inkling-Small base model.

Cognition also states that on leading evaluations SWE-2 Flash outperforms all American open-weight models, "even models that are 3x larger." The model has 276B total and 12B active parameters, the same configuration as Inkling-Small. In a generation speed demo, Cognition compares 200 tokens per second for SWE-2 Flash with 60 tokens per second for SWE-2.

The company says smaller models can increasingly handle a large share of agent work, such as automations, code review, feature development and updating documentation. SWE-2 Flash was built for that need.

How the model was trained

Cognition began with a large-scale mid-training stage on top of Inkling-Small. A central problem was choosing the data mix. The team built a held-out coding dataset meant to reflect model intelligence in real-world software engineering, then used token efficiency plots to compare mixtures. For each candidate, it measured how many tokens were needed to reach the same validation loss as a baseline run.

The final dataset consists of a few trillion tokens of scientific articles, blogs, code and other sources. Training ran in three stages so the model saw the scarcest, highest-quality data later. Cognition says held-out validation loss improved at a stable log-linear rate throughout.

Large-scale reinforcement learning followed. Cognition says it made some tweaks to the RL recipe from SWE-2, but most gains came from higher quality and diversity of RL tasks. The team audited instructions, environment setup and verifiers, and added several new task types to build general reasoning and problem-solving skills. According to the company, this improved nearly every benchmark it measured.

Cybersecurity focus and restricted weights

Cognition says enterprises and governments need observability to host and control deployments on their own infrastructure, including on-prem, and that researchers need to study how models behave and can be misused. It wants SWE-2 Flash to give industry partners and researchers transparent access to weights, with the ability to fine-tune and run the model in sensitive environments.

The company targeted cybersecurity work in training, raising the model's ExploitBench score from 11.3% to 34.8%. Because those capabilities could also help attackers, Cognition says it is not sharing the weights publicly yet. Instead it is building a program to securely share them with trusted industry partners and research labs, with more details promised "in a few weeks."

What it means for users

SWE-2 Flash is available now in Devin Desktop and CLI, and Cognition lists it as free through 11/30. For teams running routine engineering tasks, the pitch is a cheaper and faster model for everyday agent work. Organizations that need self-hosting or fine-tuning will have to wait for the weight-sharing program. The benchmark results come from Cognition itself.