Cognition released a new coding model Wednesday that costs one-fifth as much as its larger predecessor and runs roughly three times faster, but the company is keeping the finished product's weights private despite building it on an open foundation. The new model, SWE-2 Flash, was constructed on top of Thinking Machines Lab's openly available Inkling-Small, yet Cognition won't release its own version to the public, saying cybersecurity concerns drive the lockdown. The choice puts the company at odds with the broader trend of open-weight models closing the performance gap with proprietary alternatives.

Cognition says the majority of its performance improvements came from inserting a mid-training stage ahead of reinforcement learning, and the published numbers show dramatic gains. On Terminal-Bench 4.0, Flash achieved a score of 20% compared to Inkling-Small's 0.9%, while its FrontierCode 2.0 result jumped from the base model's 3.2% to 24%, slightly beating SWE-2's 23.8%. Flash nearly matched the larger model on SWE-Bench Pro v2 Hard with 89.9% versus SWE-2's 90.6%, though it fell behind on DeepSWE 1.1 and Terminal-Bench 4.0. Rival models including Claude Haiku 5.5 and DeepSeek V4.1 Flash outperformed it on multiple benchmarks. The company's own benchmark table reveals that models built outside the United States, DeepSeek V4.1 Flash among them, beat Flash on several evaluations. Flash raised its ExploitBench score from Inkling-Small's 11.3% to 34.8%, though the benchmark doesn't confirm how the model would behave in actual attacks or whether releasing its weights would boost the risk of abuse.

Cognition describes SWE-2 Flash as a fully American coding model and claims it surpasses every American open-weight model on top evaluations, including models three times its size. An AI commentary account on X, CultOfAI, characterizes mid-training as "a practical way to turn a weak open base into a usable agent model," though the same account notes the figures come from Cognition itself rather than independent verification. The company says it's creating a program to distribute the weights to selected industry partners and research labs that could examine, fine-tune, and operate the model on their own infrastructure, an arrangement likely to attract enterprises and governments concerned about sensitive code leaving their systems. More details will arrive in the coming weeks, and for now it's unclear who will meet the criteria, what limits they'll encounter, or whether they'll be permitted to distribute models built on SWE-2 Flash.

The decision to restrict access centers on cybersecurity, with the company pointing to a rise in AI-enabled cyberattacks as justification for limiting distribution. The worry is that capabilities designed to locate and examine vulnerabilities could just as readily assist attackers. Similar challenges are appearing across the industry, where AI tools can now spot tens of thousands of potential vulnerabilities in open-source projects more quickly than maintainers can investigate and patch them. Inkling-Small is a 276-billion-parameter mixture-of-experts model with 12 billion active parameters that Thinking Machines Lab, the startup headed by co-founder and CEO Mira Murati, released on July 30 under an Apache 2.0 license, meaning anyone can pull the weights from Hugging Face and build on them. Building on a model released under Apache 2.0, one of the more permissive open source licenses, doesn't require Cognition to open its own work, since the license allows developers to modify and commercialize derivatives without distributing their changes. David Hendrickson, CEO of GenerAIte Solutions, praised Cognition for transforming Inkling-Small into a "coding beast" but observed on X that "the improved weights aren't public."

Developers can access SWE-2 Flash through Devin Desktop and CLI, and the model is available at no cost through November 30. Cognition's approach of leveraging an openly licensed foundation while keeping its enhancements locked away sits within the company's legal rights but cuts against the collaborative ethos of the base it borrowed. For organizations that handle classified or proprietary code, the planned partner program could offer a middle path between fully open models and black-box APIs, though the terms remain undefined. Enterprises evaluating coding assistants now face a choice between models that promise transparency and those that prioritize controlled distribution in the name of security—a tension likely to sharpen as AI capabilities in vulnerability detection continue to advance.