What happened
Cognition introduced SWE-2 Flash as a coding model intended to lower the cost of everyday engineering work performed by software agents. It is available today in Devin Desktop and Devin CLI. Cognition says it costs one-fifth as much as its larger SWE-2 model and runs three times faster. The company’s speed comparison shows 200 tokens per second for Flash and 60 for SWE-2, using each model’s run on a Terminal-Bench 4.0 task replayed at simulated speeds. Cognition says the offer is free through November 30; the source does not specify the subsequent pricing.
The starting point is Inkling-Small, a Thinking Machines model with 276 billion total parameters and 12 billion active parameters. Cognition added a large-scale mid-training stage before reinforcement learning, a new step in its stated recipe. It says it assembled a coding dataset from a few trillion tokens of scientific articles, blogs, code and other sources, divided into three stages so that rarer, higher-quality data appeared later. It then used reinforcement learning, with changes to the task set and its quality controls compared with the SWE-2 recipe.
Cognition reports that SWE-2 Flash scored 20.0% on Terminal-Bench 4.0, compared with 0.9% for Inkling-Small, and 24.0% on FrontierCode 2.0, compared with 3.2% for the base model. It also reports a 60.1% result on DeepSWE 1.1 and 34.8% on ExploitBench, where Inkling-Small scored 17.4% and 11.3%, respectively. These figures come from the company’s own post and comparison table; the announcement does not provide an independent replication. Cognition says Flash outperforms American open-weight models on its selected evaluations, including models three times its size.
The model is not publicly downloadable. Cognition says it trained SWE-2 Flash for cybersecurity work and that the resulting capability can also be misused. It plans a program for securely sharing weights with trusted industry partners and research labs, with more details promised in a few weeks. Until then, access is through Devin’s products rather than open weights or a stated standalone API.
What the archive adds
Thinking Machines’ Inkling-Small release covered the base model’s 276B total and 12B active parameters. SWE-2 Flash is a new event built on that earlier release: Cognition reports applying mid-training and reinforcement learning for coding-agent use, while keeping weights gated. The archive also covered Cognition’s FrontierCode 2.0 benchmark; this announcement reports scores on that benchmark, but does not independently validate them.
Trend signal
This supports agent economics moving to completed work: Cognition frames a smaller model’s value around the cost and speed of engineering work in Devin, not just its parameter count. The announcement provides model-level speed and price comparisons, but no cost per verified completed task, so the link to task economics remains a direction rather than a demonstrated outcome.
Why now
Cognition says smaller models can handle a significant share of software-agent work, including code review, feature development and documentation updates. SWE-2 Flash is its response to that demand. It also describes cyber capability as a reason to limit distribution, with details of a trusted-partner and research-lab program expected in a few weeks.
Why it matters
Teams using Devin can test whether a lower-cost, faster model handles routine engineering tasks well enough to reserve larger models for harder work. The release offers a concrete comparison to try, but the company’s speed and price claims do not establish that every task will cost one-fifth as much or finish three times faster: actual results will depend on the work, agent workflow and verification burden.
The gated weights are also consequential. Teams that require on-premises or self-hosted models cannot assume that the announced partner program will meet their needs; its terms and availability are not yet specified. For now, the actionable route is to evaluate Flash inside Devin on representative work and compare completed, reviewed results—not rely on benchmark scores alone.
Limits and uncertainty
The announcement is a vendor source and does not provide independent benchmark results, detailed evaluation methodology or confidence intervals. It does not establish how SWE-2 Flash performs on a particular team’s repositories, what its post-November 30 price will be, or whether the stated speed advantage holds across tasks and configurations. The weights are not public, and Cognition has not yet described eligibility, safeguards or timing for its planned partner program. Its claim that Flash outperforms American open-weight models is limited to the evaluations and models in its own comparison.
Reader takeaway
SWE-2 Flash is available now to Devin users as a cheaper, faster option for coding-agent work, according to Cognition. Treat the published figures as a reason to run a controlled trial on representative tasks, not as a guarantee of equivalent savings or quality. Teams needing self-hosted weights should wait for the announced access program’s actual terms.