The Pulse

Cognition Releases SWE-2 Flash Coding Model in Devin Desktop and CLI, With Weights Not Yet Public

Cognition’s SWE-2 Flash is available in Devin Desktop and CLI, with faster generation and a lower price than SWE-2. The company reports gains on several coding benchmarks and is developing a program to share the model’s weights with trusted

AI.info Team · Written with AI · fact-checked · 2026-11-06

SWE-2 Flash is, Cognition says, a fully American coding model that sets a new bar for intelligence at low cost and high speed. It is available today in Devin Desktop and CLI, at one-fifth the price and three times the speed of SWE-2, the company says. Cognition is not sharing the weights publicly yet. The company is building a program to share them securely with trusted industry partners and research labs, with more information expected in a few weeks.

Reported Gains Over Inkling-Small

SWE-2 Flash improves Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0% compared with its base model, Inkling-Small, according to Cognition. On DeepSWE 1.1, the announcement lists scores of 60.1 for SWE-2 Flash and 17.4 for Inkling-Small. The new model outperforms all American open-weight models on leading evaluations, including models three times larger, the company says.

The results table also compares the two models on other tasks. SWE-2 Flash scores 45.5 on FrontierCode 1.1 Main, compared with 21.5 for Inkling-Small, and 89.2 on Terminal-Bench 2.1, compared with 64.7. On SWE-Atlas QA, their listed scores are 50.0 and 27.4; on SWE-Bench Pro v2 Hard, they are 89.9 and 56.5. SWE-2 Flash’s ExploitBench score is 34.8, up from 11.3 for Inkling-Small, Cognition reports.

Generation Listed at 200 Tokens Per Second

Cognition lists SWE-2 Flash’s generation speed at 200 tokens per second, compared with 60 tokens per second for SWE-2. Its demonstration replays each model’s run on Terminal-Bench 4.0’s gsea-proteomics task at simulated speeds of 200 and 60 tokens per second. The transcripts are shortened to equal length for the comparison. SWE-2 Flash is also free through 11/30, according to the announcement.

Training combined mid-training and reinforcement learning

Cognition added a mid-training stage before reinforcement learning, starting from Inkling-Small, which has 276 billion total parameters and 12 billion active parameters. To guide decisions about training data, the company curated a held-out coding dataset intended to reflect model intelligence in real-world software engineering scenarios. It compared candidate data mixes by measuring how many tokens each required to reach the same validation loss as a baseline mix.

The company settled on a mid-training dataset made up of a few trillion tokens of high-quality scientific articles, blogs, code and other sources. Training took place in three stages, with the scarcest, highest-quality data used later. Cognition then carried out large-scale reinforcement learning. Most of the improvements came from higher quality and greater diversity in its reinforcement-learning tasks, the company says. The company audited task instructions, environment setup and verifiers, and curated new task types to cover a wider range of capabilities.

Cognition cites cyber risks in its weights decision

Enterprises and governments need observability and the ability to host and control deployments on their own infrastructure, including on-premises, Cognition says. Researchers also need to study model behavior and potential misuse, the company says. One goal for SWE-2 Flash is to give industry partners and researchers transparent access to weights, along with the ability to fine-tune and run the model in sensitive environments.

Cognition targeted cybersecurity work during training to help defenders, raising SWE-2 Flash’s ExploitBench score from 11.3% to 34.8%, the company says. Cognition also warns that the same capabilities could be used by malicious attackers, pointing to a recent surge in cyber attacks. That risk is why the weights are not being shared publicly yet, the company says, while it develops a secure-sharing program for trusted industry partners and research labs.