SWE-2 Flash: Cognition's Faster, Lower-Cost Daily Engineering Model
Contains affiliate links; your price doesn't change · Disclosure · How we research
TL;DR
SWE-2 Flash is Cognition's coding model for everyday engineering work in Devin Desktop and CLI. Cognition says it costs about one-fifth as much as SWE-2 and runs about three times faster, with sample rates of about 200 versus 60 tokens per second. Its MoE design is labeled 276B-A12B, with about 276 billion total and 12 billion active parameters. Weights are not public. This short page keeps the vendor's relative price and speed claims rather than deriving an absolute price from them.
How this article was made: This article is research-based. Product claims come from Cognition's November 4 launch post. Devin's model documentation was also checked on November 6, 2026 at 10:25 Taipei time. We did not run the models or reproduce the reported speed. Display name: CommX.
Cognition released SWE-2 Flash for daily engineering work. Its examples include automations, code review, feature development and documentation updates.
Cognition says Flash costs about one-fifth as much as SWE-2 and runs about three times faster. Those are vendor claims, not our benchmarks.
SWE-2 Flash vs SWE-2
| Item | Official position | Practical reading |
|---|---|---|
| Relative price | About one-fifth of SWE-2 | Lower-cost tier for daily work |
| Relative speed | About 3x SWE-2 | Faster generation, according to Cognition |
| Replayed sample | About 200 tok/s versus 60 tok/s | Vendor demonstration, not an independent throughput test |
| Architecture | 276B-A12B MoE | About 276B total and 12B active parameters |
| Access | Devin Desktop and CLI | Availability depends on plan and client model settings |
| Weights | Not public | Secure sharing planned for trusted partners and research labs |
Use Flash for routine engineering and keep SWE-2 available for heavier reasoning work. That split follows Cognition's product positioning.
This page does not multiply the one-fifth claim by another price to create a Flash retail table. Check the model selector and current plan terms for your actual billing rate.
Read Cognition's SWE-2 Flash announcement
Weights and hosting are separate questions
Cognition says public weights are not available. A secure sharing program for industry partners and research labs is planned.
The API or Devin client is therefore the near-term access path for most teams. An agent application may still need separate cloud hosting for its own workers and public endpoints. See the Zeabur shutdown migration guide for that deployment decision.
For a different model with visual and Computer Use features, see Tencent Hy4.
Sources (checked November 6, 2026, Taipei)
- https://www.agentcomm.net/zh/swe-2-flash/
- https://cognition.com/blog/swe-2-flash
- https://docs.devin.ai/desktop/chat/models
FAQ
What is SWE-2 Flash?
It is Cognition's coding model for daily engineering work such as automations, code review, feature development and documentation updates. It is available in Devin Desktop and CLI.
How does SWE-2 Flash compare with SWE-2?
Cognition says Flash costs about one-fifth as much and generates about three times faster. Its replayed sample uses about 200 tokens per second for Flash and 60 for SWE-2. These are vendor figures.
What does 276B-A12B mean?
It describes a mixture-of-experts model with about 276 billion total parameters and 12 billion active parameters. It is not a verified local hardware requirement.
Are SWE-2 Flash weights public?
No. Cognition says it is building a secure sharing program for trusted industry partners and research labs instead of publishing the weights broadly.
Key takeaways
- SWE-2 Flash is Cognition's coding model for everyday engineering work in Devin Desktop and CLI.
- Cognition says it costs about one-fifth as much as SWE-2 and runs about three times faster, with sample rates of about 200 versus 60 tokens per second.
- Its MoE design is labeled 276B-A12B, with about 276 billion total and 12 billion active parameters.
SourcesRead Cognition's SWE-2 Flash announcement · docs.devin.ai/desktop/chat
Last updated
Cite this pageSWE-2 Flash: Cognition's Faster, Lower-Cost Daily Engineering Model — https://www.agentcomm.net/articles/swe-2-flash/