On Wednesday, Cognition unveiled SWE-2 Flash, a coding model that costs one-fifth the price of its larger SWE-2 counterpart and processes tokens roughly three times faster. The foundation for this advancement comes from Thinking Machines Lab's Inkling-Small, an open-weight model that anyone can download and modify. However, Cognition has chosen to keep the weights of its own finished model under lock, citing cybersecurity concerns as the reason for the restriction.

Inkling-Small, developed by Thinking Machines Lab under the leadership of CEO Mira Murati, is a 276-billion-parameter mixture-of-experts architecture with 12 billion active parameters. The company released it on July 30 under an Apache 2.0 license, making it freely available on Hugging Face for anyone to download and build upon.

According to Cognition, the performance improvements stem primarily from introducing a mid-training stage that precedes reinforcement learning. The results are substantial: on Terminal-Bench 4.0, Flash achieved 20% compared to Inkling-Small's 0.9%, and on FrontierCode 2.0, it jumped from the base model's 3.2% to 24%, slightly surpassing SWE-2's 23.8%. CultOfAI, an AI commentary account on X, describes this approach as "a practical way to turn a weak open base into a usable agent model," though it notes that these figures come from Cognition itself rather than independent verification.

Performance trade-offs and competitive standing

Flash does not consistently outperform the larger model across all benchmarks. On SWE-Bench Pro v2 Hard, it scores 89.9% to SWE-2's 90.6%, a narrow gap, but it lags on DeepSWE 1.1 and Terminal-Bench 4.0. Competitors such as Claude Haiku 5.5 and DeepSeek V4.1 Flash surpass it on several evaluations.

Open foundation, proprietary outcome

The Apache 2.0 license governing Inkling-Small permits developers to modify and commercialize derivatives without disclosing their changes. This means Cognition's decision to keep SWE-2 Flash's weights private is legally permissible, even though it diverges from the collaborative ethos of the underlying open-source project.

David Hendrickson, CEO of GenerAIte Solutions, acknowledged Cognition's success in transforming Inkling-Small into a "coding beast" but highlighted on X that "the improved weights aren't public." His observation reflects a broader shift in the field: open-weight models are narrowing the performance gap with proprietary systems and emerging as viable alternatives for coding platforms. Cognition's choice to restrict access runs counter to this trend.

Cognition is developing a program to share weights with selected industry partners and research institutions, allowing them to inspect, fine-tune, and deploy the model within their own infrastructure. This arrangement may appeal to enterprises and government agencies concerned about sensitive code leaving their premises. The company has indicated that additional details will emerge in the coming weeks, though it remains unclear which organizations will qualify, what limitations they will encounter, or whether they can distribute models derived from SWE-2 Flash.

Security rationale for restricted access

Cybersecurity considerations form the cornerstone of Cognition's decision to restrict the weights. Flash's training included cybersecurity tasks, raising its ExploitBench score from Inkling-Small's 11.3% to 34.8%. While this represents meaningful progress, the benchmark does not demonstrate how the model would behave in actual attack scenarios or whether releasing the weights would amplify misuse risks.

Cognition points to a rise in AI-enabled cyberattacks as justification for limiting access. The underlying concern is that capabilities designed to identify and investigate vulnerabilities could equally serve attackers. This challenge extends across the industry: AI systems can now detect tens of thousands of potential vulnerabilities in open-source projects faster than maintainers can patch them.

American model, limited reach

Cognition characterizes SWE-2 Flash as a fully American coding model and asserts that it outperforms every American open-weight model on major benchmarks, including models three times its size. This distinction carries significance because Cognition's own benchmark table reveals that models developed outside the United States, including DeepSeek V4.1 Flash, achieve higher scores on several tests.

Developers can access SWE-2 Flash through Devin Desktop and CLI, with free availability extending through November 30.