Cognition adds a lower-priced coding model to Devin, keeping its weights private

Cognition released SWE-2 Flash on November 4, 2026, giving Devin users a smaller coding model aimed at everyday engineering work. Available in Devin Desktop and its command-line interface, the release pairs claimed speed and price advantages over SWE-2 with a restriction: Cognition is not making the trained model’s weights publicly downloadable.

A cheaper sibling, not a clean sweep

Cognition positions Flash for automations, code review, feature development and documentation updates. The company says it runs at three times the speed and one-fifth the price of SWE-2 inside Devin. It is also advertising free use through November 30, a temporary offer distinct from that relative price comparison.

The speed illustration needs a careful reading. Cognition shows transcripts from each model’s run on the same Terminal-Bench 4.0 task, replayed at a simulated 200 tokens per second for Flash and 60 for SWE-2. The transcripts are shortened to equal length. That illustrates the stated output-speed difference; it is not a measurement of how long each complete task took.

Cognition’s benchmark table also resists a simple winner label. Flash scores 24.0% on FrontierCode 2.0, narrowly ahead of SWE-2’s 23.8%. On Terminal-Bench 4.0, Flash trails at 20.0% against 27.3%. The lower-priced model therefore does not consistently match or exceed its larger sibling, even within the company’s own evaluations.

Nearly level on one coding evaluation

0189.9%

SWE-2 Flash

Cognition reports this score on SWE-Bench Pro v2 Hard for its smaller model.

0290.6%

SWE-2

The larger sibling scores slightly higher on the same evaluation in Cognition’s table.

The training changed more than the model size

Flash starts from Thinking Machines’ Inkling-Small, with 276 billion total parameters and 12 billion active parameters. Parameters are the numerical values that encode learned behavior. Inkling-Small uses a mixture-of-experts design: each token is routed through selected specialist components, alongside shared components, rather than activating the entire model.

For the first time in its training recipe, Cognition added a mid-training stage before reinforcement learning. It assembled a few trillion tokens of scientific articles, blogs, code and other material. To choose the mixture, the team compared how many tokens different datasets needed to reach the same validation loss—a measure of prediction error on held-out data—as a baseline.

That phase ran in three stages, reserving the scarcest, highest-quality data for later. Cognition then applied large-scale reinforcement learning, training through task feedback. It attributes most improvements in that phase to better and more varied tasks, including audited instructions, environments and automated checks, rather than tweaks to its earlier reinforcement-learning recipe.

The reported gains over the starting model are substantial: Terminal-Bench 4.0 rises from 0.9% to 20.0%, while FrontierCode 2.0 increases from 3.2% to 24.0%. Those comparisons describe the result of Cognition’s added training as a whole; they do not isolate the contribution of each stage.

An open base, a restricted derivative

Thinking Machines’ model card lists Inkling-Small under Apache 2.0 and provides downloadable weights, the learned numerical values needed to run the model. That access applies to the base model. Cognition is withholding the trained SWE-2 Flash weights, so the base model’s availability should not be mistaken for a downloadable Flash release.

Cognition cites cybersecurity risk for that decision. It says targeted security training raised the model’s ExploitBench score from 11.3% to 34.8%. The company wants those capabilities to help defenders, but acknowledges that malicious attackers could use them too. Its response is selective distribution rather than a public weights release.

The planned weights-sharing program would serve trusted industry partners and research labs. Cognition describes the goal as enabling inspection, fine-tuning and deployments in sensitive environments, including on-premises infrastructure. Those are aims for the program being built, not general access delivered with this launch. The company promises more details in a few weeks; today’s confirmed product access is through Devin Desktop and CLI.