Cognition released SWE-2 Flash on 4 November, giving users of its Devin coding agent a smaller model for routine software work. The company says Flash runs three times faster than SWE-2 at one-fifth the price, Startup Fortune reported. Flash is available in Devin Desktop and Devin CLI, where Cognition is positioning it for tasks such as code review, documentation updates and routine feature development.
Key points
- Flash is based on Thinking Machines Lab’s Inkling-Small and activates 12 billion of its 276 billion parameters at inference.
- Cognition claims three times the speed and one-fifth the price of SWE-2, but its speed illustration replays runs at simulated rates.
- Cognition’s benchmark scores favour Flash on FrontierCode 2.0 and SWE-2 on Terminal-Bench 4.0 and DeepSWE 1.1.
- Cognition is withholding Flash’s weights from public release, citing cybersecurity risks.
Inkling-Small changes the size of SWE-2
Flash is built on Inkling-Small, a model from Thinking Machines Lab, the company founded by Mira Murati. It has 276 billion parameters in total, with 12 billion active when generating a token, Unite.AI reported. Thinking Machines released Inkling-Small’s weights on 30 July 2026 under the Apache 2.0 licence. Cognition adapted that base for its coding model rather than publishing the base model unchanged.
The distinction between total and active parameters matters because Inkling-Small uses a mixture-of-experts design. Think of a workshop with many specialised benches: the whole workshop is available, but a particular job goes to only some of them. For each token it produces, the model uses a fraction of its stored parameters rather than running all 276 billion. SWE-2, released on 10 September, was instead built on Moonshot AI’s Kimi K3, which has 2.8 trillion parameters in total and 104 billion active per token, OrcaRouter reported.
Updating documentation could require several small checks against the code it describes. A model using fewer active parameters for each response could make that sequence less computationally demanding, if its responses were adequate for the work. That is the practical attraction of a smaller model inside an agent that repeatedly reads, writes and checks files, rather than answering one request and stopping.
Cognition also changed the training process. It says it placed a substantial mid-training stage before reinforcement learning, selected its data mixture using a held-out coding dataset, and then trained on a few trillion tokens from code, scientific articles, blogs and other material. Candidate mixtures were judged by how many tokens they needed to reach a baseline validation loss. Cognition attributes most of Flash’s improvement over its earlier training recipe to better and more varied reinforcement-learning tasks, rather than adjustments to the recipe itself.
Terminal-Bench 4.0 separates speed from scores
Cognition’s three-times-faster claim comes with a particular method. Each model’s run on a Terminal-Bench 4.0 task was replayed at a simulated rate of 200 tokens per second for Flash and 60 for SWE-2, with the transcripts shortened to equal length, OrcaRouter reported. Those rates were inputs to the replay, rather than timings of the two models serving requests under production load. The one-fifth price comparison concerns use inside Devin, where an agent’s work includes more than the model generating tokens.
Cognition’s scores, reported by Unite.AI, give a mixed picture across the coding evaluations. Flash scored 24.0 on FrontierCode 2.0 against SWE-2’s 23.8. On Terminal-Bench 4.0, Flash scored 20.0 against 27.3, while on DeepSWE 1.1 it scored 60.1 against 73.0. These are Cognition’s evaluations of the models on named tests, with the larger SWE-2 ahead on two of the three reported here. In a separate version of Terminal-Bench, Cognition listed Flash at 89.2 on Terminal-Bench 2.1, FourWeekMBA reported.
Cognition restricts access to Flash’s weights
Access through Devin does not include access to Flash’s weights, the numerical settings needed to run or adapt the model independently. Cognition is withholding them from public release because it says the model’s cybersecurity capabilities could be misused. In Cognition’s own testing, targeted cybersecurity training raised its ExploitBench score from Inkling-Small’s 11.3% to 34.8%, Unite.AI reported. The company is developing a secure sharing programme for partners and researchers that it says could permit fine-tuning and deployment in customer-controlled environments.
For developers building outside Devin, that makes the distribution choice consequential: Inkling-Small’s weights are available under Apache 2.0, while Cognition’s coding adaptation is currently accessible through its agent products. Cognition is offering Flash free through 30 November for Pro, Max and Teams users in Devin Desktop and Devin CLI, Startup Fortune reported.
