SWE-2 Flash

SWE-2 Flash is Cognition’s fast, lower-cost coding model, launched 2026-11-04 inside Devin. Cognition positions it as the speed/efficiency tier of the SWE-2 line, claiming it outperforms all American open-weight models on leading evals and stressing a fully U.S. stack. Community posts describe ~276B / 12B active (built on Thinking Machines Inkling-Small) and ~200 TPS, available in Devin Desktop/CLI and free through 2026-11-30.

Versus full SWE-2, Flash targets routine automation, review, feature work, and docs—agent jobs where latency and cost dominate—leaving heavier lifts to higher-tier models.

Problem

Devin users need a fast/cheap daily coding tier without always paying frontier rates. SWE-2 Flash is that speed tier in the SWE-2 family.

Differentiation

Closed model inside Devin—not open weights. Aimed at beating U.S. open coders on speed/cost, not at being a general chatbot.

Limits

• Requires Devin access; no public weight drop.
• Benchmarks are vendor/community-reported—validate on your repos.
• Pricing after the free window follows Cognition’s plans.

Who it’s for

Devin teams that want a faster/cheaper model for routine PRs, review, and scaffolding.

Get started

1. Read the Cognition post.
2. Select SWE-2 Flash in Devin Desktop/CLI (free through Nov 30).
3. A/B a real backlog against SWE-2 and other models on time and rework.