AI
Cognition Boosts Open Model into Powerful Coder—Then Keeps the Improved Weights Secret
The company's new SWE-2 Flash, built on an open-weight foundation, delivers strong performance but remains proprietary, citing security risks.
1/2
The New Stack
Cognition has significantly enhanced an openly available coding model but is withholding the refined version from public release. The newly launched SWE-2 Flash operates at one-fifth the cost of Cognition's larger SWE-2 offering while running roughly three times faster, yet the company is keeping its trained weights behind closed doors over cybersecurity considerations.
How the Performance Leap Happened#
SWE-2 Flash is built on Inkling-Small, a 276-billion-parameter open-weight model from Thinking Machines Lab (led by Mira Murati) that became available in July under an Apache 2.0 license. Cognition's development process centered on inserting a dedicated mid-training phase prior to reinforcement learning—an approach some observers describe as "a practical way to turn a weak open base into a usable agent model."
The results are striking on several fronts. Flash achieved 20% on Terminal-Bench 4.0 versus Inkling-Small's 0.9%, and jumped to 24% on FrontierCode 2.0 (up from 3.2%), narrowly outpacing SWE-2's 23.8%. However, independent verification of these figures remains unavailable; they reflect Cognition's own measurements.
Where Flash Stands Against Rivals#
The smaller model keeps pace with SWE-2 on some tests—reaching 89.9% versus SWE-2's 90.6% on SWE-Bench Pro v2 Hard—but trails on others like DeepSWE 1.1 and Terminal-Bench 4.0. Meanwhile, competing options including Claude Haiku 5.5 and DeepSeek V4.1 Flash surpass Flash on multiple benchmarks. Despite these gaps, Cognition labels it the top-performing American open-weight coding model, a claim that becomes qualified when considering that non-U.S. alternatives score higher in several categories.
The Paradox: Built on Open, Released Closed#
Starting with a model published under one of open source's most permissive licenses doesn't obligate Cognition to release its own derivative. The Apache 2.0 license allows modification and commercial use without requiring disclosure of improvements—so the decision to keep weights proprietary falls within legal bounds. Yet it diverges sharply from the collaborative spirit underlying the foundation it relied upon.
The move drew notice from David Hendrickson, CEO of GenerAIte Solutions, who acknowledged Cognition's achievement in transforming Inkling-Small into a while highlighting that As open-weight models increasingly close the performance gap with proprietary systems and stake stronger claims in coding platforms, Cognition's approach stands out as atypical.

