Cognition Launches SWE-2 Flash, a Code Model Built on Inkling-Small with Withheld Weights

Cognition, the developer of the Devin software engineering agent, introduced SWE-2 Flash on November 4, 2026—a code model designed for routine agent workflows: code reviews, automations, small features, and documentation updates. Built on Inkling-Small, the open-weight model from Thinking Machines, Cognition describes it as an “all-American” model. But unlike its foundation, it cannot be downloaded: the company is withholding the weights, citing cybersecurity risks.

An Open Foundation, Transformed by Training

Inkling-Small is a Mixture-of-Experts model with 276 billion parameters, of which only 12 billion are activated per token. According to its Hugging Face model card, it features 42 layers, routes each token to 6 experts out of 256 (plus 2 shared experts), and accepts text, image, and audio inputs. Its weights are released under an Apache 2.0 license, which permits this kind of derivative work without requiring the result to be open-sourced.

Cognition says it introduced, for the first time in its training pipeline, an intermediate mid-training phase prior to reinforcement learning: several trillion tokens of scientific papers, blog posts, code, and other sources, structured across three stages so that the model encounters the rarest and highest-quality data toward the end. This was followed by a reinforcement learning phase, with gains primarily driven, according to Cognition, by more diverse and rigorously verified tasks than those used for SWE-2.

The claimed impact is dramatic on certain benchmarks: on Terminal-Bench 4.0, the score jumps from 0.9% for Inkling-Small to 20.0%; on FrontierCode 2.0, from 3.2 to 24.0. Cognition asserts that SWE-2 Flash outperforms every American open-weight model in its comparison table, including models three times its size such as Beam 1.0 or Muse Spark Open.

Smaller Than SWE-2, Not Universally Superior

Cognition’s benchmark table also highlights the model’s limitations. Compared to its flagship counterpart SWE-2 (2,800 billion parameters, 104 billion active), SWE-2 Flash is slightly ahead on FrontierCode 2.0 (24.0 versus 23.8) but trails substantially on Terminal-Bench 4.0 (20.0 versus 27.3) and DeepSWE 1.1 (60.1 versus 73.0). Several models outside the “American open-weight” category fare better on certain evaluations: DeepSeek V4.1 Flash hits 42.3 on Terminal-Bench 4.0 and 77.0 on DeepSWE 1.1 in the same chart, while Claude Haiku 5.5 leads on FrontierCode 2.0 (27.7).

In terms of deployment, SWE-2 Flash is available in Devin Desktop and the Devin CLI, free of charge through the end of November for Pro, Max, and Teams subscribers. Cognition frames it as five times cheaper and three times faster than SWE-2, but has not released per-token pricing or an independent API outside of Devin. The blog post’s speed demonstration (200 versus 60 tokens per second) replays real sessions at simulated throughputs, as noted in its caption: it is not a benchmark of live production performance.

Weights Withheld Citing Cybersecurity Risks

The most debated aspect of the release is its distribution. Cognition reports having explicitly trained the model on cybersecurity tasks “to assist defenders,” boosting its score on ExploitBench from 11.3% to 34.8%. Because these capabilities could equally be leveraged by attackers, the company is not releasing the weights for now. Instead, it is planning a secure access program for industry partners and research labs that wish to fine-tune and run the model within their own environments, promising more details “in the coming weeks.”