On November 4, 2026, Cognition introduced SWE-2 Flash, a coding model aimed at everyday engineering work inside Devin Desktop and Devin CLI. Cognition says it is about one-fifth the price and three times the speed of its larger sibling, SWE-2. The weights are American-trained but not public.

Coffee summary

  • FACT: SWE-2 Flash is built on Thinking Machines’ Inkling-Small (276B total / 12B active). Cognition added a mid-training stage before reinforcement learning.
  • CLAIM (Cognition): Terminal-Bench 4.0 rose from 0.9 to 20.0 and FrontierCode 2.0 from 3.2 to 24.0 versus the Inkling-Small base; Cognition says it beats US open-weight models even 3× larger on leading evals.
  • FACT (Devin docs): SWE-2 Flash is free in Devin Desktop and CLI until December 4, 2026 at 18:00 UTC; Medium is on Free plans, High/Max need paid plans.
  • FACT: Weights are not public. Cognition cites cyber risk (ExploitBench 11.3→34.8, vendor figure) and says a secure share program for partners is coming.
  • CONTEXT: The on-page speed demo replays a Terminal-Bench task at a simulated 200 vs 60 tokens/sec — it is not a live timed race.

What happened

Cognition’s recipe change is the mid-training stage: a few trillion tokens of curated scientific articles, blogs, code, and other sources, tuned with held-out coding validation and token-efficiency plots, then large-scale RL on audited, more diverse tasks. The company positions Flash for automations, code review, features, and docs — the bulk of agent work — while reserving SWE-2 for heavier jobs.

Availability is product-first. Devin docs confirm Flash is in Desktop and CLI, not Devin cloud, with a clear free window through early December, then list token rates afterward.

Why it matters

Coding agents are shifting from “one giant model for everything” to a mix of fast cheap models and occasional frontier calls. If Cognition’s cost and speed claims hold in real repos, teams can keep Devin on for routine work without burning the SWE-2 budget. The withheld weights also show the new tension: useful cyber capability is exactly why labs hesitate to open models fully.

Who should care

  • Devin Desktop/CLI users who can try Flash free through December 4.
  • Engineering leads balancing agent spend against review quality.
  • Security-conscious orgs that want on-prem weights later via Cognition’s planned share program.

Limitations

All benchmark numbers are Cognition’s. Independent replications are not in the launch post. ExploitBench gains cut both ways: helpful for defenders, risky if weights leak. The speed comparison on the blog is an equal-length simulated replay. Flash is not in Devin cloud sessions.

What to do next

  1. In Devin Desktop or CLI, select SWE-2 Flash Medium on a Free plan (or High/Max if paid) before December 4 while it is free.
  2. Run the same small review or docs task on Flash and on your usual model; compare edits accepted, not just tokens/sec.
  3. Do not expect a public Hugging Face dump; watch Cognition’s follow-up on the secure weight-sharing program.

AIImpish Take

SWE-2 Flash is Cognition optimizing for the middle of the workload, with vendor scores that look strong and a free month to verify them yourself. Treat the cyber score as a warning label as much as a feature, and judge the model on your diffs.

Related on AIImpish

Sources: Cognition, Introducing SWE-2 Flash; Devin docs, AI Models pricing and free window.