All news

Cognition releases SWE-2 Flash, three times faster and one-fifth the price of SWE-2

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

A cheaper, faster model makes frequent agentic coding tasks more accessible, while restricted weights preserve control over its cyber capabilities.

SWE-2 Flash is based on Thinking Machines' Inkling-Small with 276 billion total and 12 billion active parameters. Cognition added a mid-training stage before reinforcement learning for the first time. According to the company, Terminal-Bench 4.0 improved from 0.9% for the base model to 20%, while FrontierCode 2.0 rose from 3.2% to 24%. A demonstration reached 200 tokens per second compared with 60 for SWE-2. The model is available in Devin Desktop and CLI and is free to users through 30 November. After its ExploitBench score rose from 11.3% to 34.8%, Cognition decided not to publish the weights openly. It is preparing controlled access for vetted industry partners and research laboratories. All quality and speed comparisons come from Cognition.

History

Cognition develops the SWE-2 coding models

2 events
  1. Cognition releases SWE-2 Flash, three times faster and one-fifth the price of SWE-2Open now
  2. Cognition releases coding model SWE-2 built on China's Kimi K3 — nearly on par with GPT-6 Astra at a quarter of the price