Cognition releases SWE-2 Flash coding model post-trained from Inkling 2 Small
Matthew Berman calls Cognition's SWE-2 Flash, a post-trained version of Inkling 2 Small, a fast model meant for coding sub-agents.
Cognition has released SWE-2 Flash, a coding model it built by post-training Thinking Machines' Inkling-Small. The weights are not being published. Cognition's post says "we are not sharing the weights publicly yet," citing the cyber risks of a capable model and a recent surge in cyber attacks. Instead, it says it is building a program to share the weights with trusted industry partners and research labs, with "more on this in a few weeks." Cognition calls the model a "fully American coding model."
Inkling-Small is a model with 276B total parameters and 12B active. Before reinforcement learning, Cognition ran a large mid-training stage on a few trillion tokens of scientific articles, blogs, code and other sources. The stage was split into three parts so the scarcest, highest-quality data came last. To choose the mixture, the team measured how many tokens each candidate mix needed to reach the same validation loss as a baseline, plotting that ratio as a token-efficiency measure. Cognition says the reinforcement learning stage made some tweaks to the recipe from its earlier SWE-2 model. Most of the gains, it says, came from auditing the instructions, environments and verifiers for its tasks and adding new task types.
Headline numbers show large jumps from a small starting point. Cognition reports Terminal-Bench 4.0 rising from 0.9% for Inkling-Small to 20.0%. FrontierCode 2.0 rises from 3.2% to 24.0%, and ExploitBench rises from 11.3% to 34.8%, the last reflecting work targeted at cybersecurity. Cognition also claims SWE-2 Flash "outperforms all American open-weight models" on leading evals, including models three times larger. Those comparisons come from Cognition's own tables and its chosen set of competitors. The post does not describe the evaluation setup beyond the benchmark names.
The pitch centers on cost and speed. Cognition says SWE-2 Flash is available now in Devin Desktop and its CLI, at one-fifth the price and three times the speed of its larger sibling, SWE-2. It is free through 11/30. The speed comparison replays one Terminal-Bench 4.0 task, simulated at 200 tokens per second for SWE-2 Flash and 60 for SWE-2, with transcripts shortened to equal length. That illustrates throughput, but it is one task replayed at simulated rates, not a measured benchmark across workloads.
Matthew Berman described the model as a fast, efficient coding model meant to be delegated to in coding tasks. In his framing, a larger model would do the planning and hand work to it as a sub-agent. He presented it as one more example of labs building on open-weight models, and gave no benchmark figures for it. Cognition's post does not use the sub-agent framing. It says smaller models are increasingly able to handle automations, code review, feature development and documentation updates. That fits the delegation idea but is a narrower claim.
What changed
Cognition released SWE-2 Flash, a coding model post-trained from Inkling-Small, available now in Devin Desktop and its CLI, with weights not published.
Type: New capability · Applies to: SWE-2 Flash, Inkling-Small, SWE-2, Devin Desktop, CLI
Why it matters
Coding tools may route routine sub-tasks to smaller, faster models, which can lower costs for teams running agent workflows.
Original source
Where we saw it
Forward Future (Matthew Berman)
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works