Cognition has introduced SWE-2 Flash, a smaller, faster coding model built for the everyday work software agents actually do. The model runs at three times the speed and one-fifth the price of its larger sibling, SWE-2, and is available now in Devin Desktop and the Devin command-line tool. Cognition is offering it free through November 30. The launch, announced on November 4, targets routine engineering rather than benchmark leadership.
What Flash is
Flash is built on Inkling-Small, a 276-billion-parameter model from Thinking Machines Lab, with 12 billion active parameters per token. Cognition added a mid-training stage before reinforcement learning, training on a few trillion tokens of articles, blogs, and code. The result lifts Terminal-Bench 4.0 from 0.9 percent to 20.0 percent and FrontierCode 2.0 from 3.2 percent to 24.0 percent.
Cognition says Flash outperforms all American open-weight models on the evaluations it reports, even models three times larger, though it stops short of claiming the overall lead. The pitch is deliberately narrow: smaller models can now handle much agent work, including automations, code review, feature development, and documentation updates, where speed and price matter more than peak capability.
The weights decision
The most telling choice is what Cognition is not doing: it is not publishing the weights. Targeted cybersecurity training lifted Flash’s ExploitBench score from 11.3 percent to 34.8 percent, and Cognition says those same capabilities could help malicious attackers, pointing to a recent surge in cyber attacks.
Instead of an open release, Cognition is building a program to share the weights securely with trusted industry partners and research labs, with details promised in a few weeks. Enterprises and governments that need to host models on their own infrastructure get a path to transparency; researchers get a way to study the model without a public download.
Our read
Our read: Flash is Cognition admitting that the coding-agent race has two contests, not one. The flagship contest chases the hardest benchmarks, while the volume contest decides which model handles tickets, reviews, and refactors. A model at one-fifth the price and three times the speed wins the second contest while losing the first, and that is where the money is.
The same week showed how crowded that race has become, with OpenAI buying Pierre for agent version control and Meta buying Conductor for parallel coding agents. Cognition’s answer is different: instead of buying distribution, it is cutting the cost of the work itself. If Flash holds up in daily use, rivals will have to answer on price.
Why It Matters
For engineering teams, Flash changes the arithmetic of putting agents on work. Reviewing pull requests, triaging issues, and updating documentation are jobs where a slower, pricier flagship is overkill, and a fast, cheap model can run on every ticket instead of a few. Free use through November 30 gives teams a no-cost window to test the trade on their code.
For Cognition, the launch also tests a safety argument the industry is watching: that powerful models can be shared without being published. If the secure-sharing program satisfies partners and researchers, other labs with dual-use worries may copy it. If it feels like a velvet rope, the withheld weights will read as marketing, and the open-weight field will keep its scepticism.
Background
Background: Flash follows SWE-2, Cognition’s most capable coding model, which launched on September 10 and was built by post-training Moonshot’s Kimi K3. The new model takes the opposite approach, starting from an American open-weight base and trading peak scores for speed and cost. Google’s Antigravity agent shows the same pattern elsewhere, hunting for new drugs with specialised skills.
Source: Cognition’s SWE-2 Flash announcement, Unite.AI’s launch report
0 Comments