Cognition withholds its SWE-2 Flash weights over cyber risk
Built on Thinking Machines' open Inkling-Small, the coding model's Terminal-Bench 4.0 score rose from 0.9% to 20.0% after Cognition's training, by the company's account; weights go only to vetted partners.
- Models & capabilities
- Security & misuse
- Minor
Cognition released SWE-2 Flash, a coding model available in its Devin Desktop and command-line products, and said it would not publish the weights. It is built on Inkling-Small, the open-weight model from Thinking Machines with 276 billion total and 12 billion active parameters. Cognition said the weights are not being shared publicly “because of cyber risks”, and that it plans to share them securely with trusted industry partners and research labs through a programme it will describe in a few weeks.
The company described the training in two stages: a mid-training run on a curated mix of a few trillion tokens of scientific articles, blogs, code and other sources, followed by large-scale reinforcement learning. It reported that Terminal-Bench 4.0 rose from 0.9% for the base model to 20.0%, and FrontierCode 2.0 from 3.2% to 24.0%. In its comparison, Inkling-Small scored 4.0 on Terminal-Bench, and another model it listed, Muse Spark Open, scored 27.3. These are Cognition’s own figures. It also said the model outperforms all American open-weight models on leading evaluations, including some about three times its size. SWE-2 Flash is priced at one-fifth the cost and three times the speed of SWE-2, a larger model listed in the same tables at 2.8 trillion total parameters and 104 billion active, and is free until 30 November.
The cyber reasoning is Cognition’s own. It said it targeted cybersecurity work during training, raising its ExploitBench score from 11.3% to 34.8%, and cautioned that “These same capabilities can also be used by malicious attackers”. The post does not describe any independent evaluation of the model’s offensive capability.
The decision places Cognition alongside labs now restricting weights for cyber reasons, though by a different route from Moonshot’s approach with Kimi K3.1 Pro, which removed cyber data and then published the weights. It also shows an open-weight model being used as raw material for a closed derivative: the following day Thinking Machines announced a successor, Inkling 2 Small, with weights to follow after external safety testing.