OPEN LINK ↗
// 1h agoMODEL RELEASE
Cognition SWE-2 Flash Trades Weights for Speed
Cognition’s SWE-2 Flash is a fast coding model built from Thinking Machines’ Inkling-Small, with reported performance above American open-weight models up to three times larger. It runs in Devin Desktop and CLI, but Cognition is withholding public weights while developing a secure sharing program.
// ANALYSIS
SWE-2 Flash is an efficiency play: prove a smaller active model can deliver strong coding results, then keep distribution tightly controlled.
- –Cognition reports roughly three-times faster inference and one-fifth the cost of SWE-2
- –Built from Inkling-Small with 276B total and 12B active parameters, plus additional mid-training and reinforcement learning
- –Company benchmarks show major gains over the base model, including Terminal-Bench 4.0 and ExploitBench
- –The model does not lead every evaluation, so broad “best open-weight” claims need careful benchmark-by-benchmark scrutiny
- –Free access for paid Devin users through November 30 gives developers a low-friction window to test it
// TAGS
swe-2-flashllmai-codingcoding-agentopen-weightsevaluationsecurity
DISCOVERED
1h ago
2026-11-04
PUBLISHED
2h ago
2026-11-04
RELEVANCE
9/ 10
AUTHOR
cognition