On November 4 2026, Cognition, the team behind the Devin coding agent, released SWE-2 Flash, a US-developed open-weight code model built to be cheap and fast. With 276 billion total parameters and only 12 billion active per inference, it generates at 200 tokens per second, three times the speed of its flagship SWE-2, at one-fifth the price. Cognition says it outperforms every other US open-weight model.
Specs and Speed
SWE-2 Flash is a sparse mixture-of-experts model built on Thinking Machines' Inkling-Small base. Against the flagship SWE-2, which carries 2.8 trillion total parameters and 104 billion active, the Flash variant is far smaller but far more efficient. Native generation runs at 200 tokens per second versus 60 for SWE-2, a threefold speedup, while pricing lands at one-fifth of the flagship rate.
Where It Runs and Who It Is For
The model is already live inside the Devin desktop client and Devin command-line tool, targeting high-frequency software-engineering work: code review, feature iteration, documentation upkeep and terminal operations. Cognition opened a free trial window that closes on November 30 2026.
Why It Matters
Cognition positions SWE-2 Flash as the strongest US open-weight code model, beating competitors with up to three times its parameter count. In a market where frontier coding agents are gated behind closed APIs and high bills, a fast, openly licensed, domestically developed alternative changes the cost math for teams shipping software with agents.
Key Points
- 276 billion total and 12 billion active parameters (sparse MoE)
- 200 tokens per second generation, three times faster than SWE-2 at 60 tokens per second
- Priced at one-fifth of the flagship SWE-2
- Free trial window closes November 30 2026
- Built on Thinking Machines' Inkling-Small base
Full briefing: https://www.threerouter.com/blog/326
Sources: Tonghuashun Finance, Cognition, threerouter
Top comments (0)