THE ESSENTIALS IN 30 SECONDS

  • SWE-2 Flash is built on Thinking Machines' Inkling-Small, a mixture-of-experts model with 276 billion total and 12 billion active parameters.
  • Cognition says it costs one-fifth as much as SWE-2 and runs about three times as fast; it is free in Devin through November 30.
  • The weights are not public, which Cognition attributes to cyber risks; a gated sharing programme for partners and labs is in preparation.

Cognition, the company behind the Devin coding agent, released SWE-2 Flash on November 4. It is a smaller, cheaper sibling of SWE-2, which the company launched in September, and is aimed at routine engineering work. The model is available now in Devin Desktop and the Devin CLI, and Cognition is offering it free of charge until November 30.

What the model is

SWE-2 Flash is built on Inkling-Small, an American open-weight base model from Thinking Machines with 276 billion total parameters, of which 12 billion are active per token. SWE-2, by comparison, lists 2.8 trillion total and 104 billion active parameters. Cognition says it added a mid-training stage before reinforcement learning for the first time, and that it trained specifically for cybersecurity tasks. The company describes the result as "a fully American coding model".

On price and speed, Cognition claims SWE-2 Flash costs one-fifth as much as SWE-2 and is three times faster. The speed figure is illustrated by a replay of a single Terminal-Bench 4.0 task, generating about 200 tokens per second against 60 for SWE-2, rather than by a broad measurement. The blog post does not give per-token prices.

How it compares

In Cognition's own benchmark table, SWE-2 Flash scores 24.0% on FrontierCode 2.0 and 20.0% on Terminal-Bench 4.0, up from 3.2% and 0.9% for the untuned Inkling-Small. On ExploitBench it rises from 11.3% to 34.8%. Among the open-weight models in the table, it leads on FrontierCode 2.0, but it trails several of them elsewhere: GLM 5.3 Flash and DeepSeek V4.1 Flash, for example, post about 42% on Terminal-Bench 4.0. SWE-2 remains ahead on most tests. These are vendor figures that have not been independently reproduced.

Weights stay private for now

Although the base model is open-weight, Cognition is not publishing SWE-2 Flash's weights, citing cyber risks. It says it is building a programme to share them securely with trusted industry partners and research labs, so they can fine-tune the model and run it in sensitive environments, including on premises, with details promised in the coming weeks. RuntimeWire notes that this keeps distribution under Cognition's control. Pricing after the free period has not been announced.