Is SWE-2 Flash state of the art?Pending benchmarks.Announced Nov 4, 2026; no independent results yet.
Cognition's smaller coding model, mid-trained and RL-trained from Thinking Machines' Inkling-Small (276B total, 12B active parameters). Available inside Devin; weights are not public, with a restricted weight-sharing program announced.
The details so far
- Lab
- Cognition
- Announced
- Nov 4, 2026
- API
- Not yet
- Open weights
- Announced, not yet published · Not public; Cognition is building a program to share weights with vetted industry partners and researchers
- Parameters
- 276B total, 12B active (Inkling-Small base)
- Base model
- Thinking Machines Inkling-Small (276B total, 12B active)
- Training
- Mid-training on a few trillion tokens in three stages, then large-scale RL
- Access
- In Devin (Devin Desktop and Devin CLI); free for Pro, Max and Teams users until the end of November 2026
- Sibling
- Separate from SWE-2, Cognition's larger coding model
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: one-fifth the price and 3x the speed of SWE-2.
Lab claim: improves on Inkling-Small from 0.9% to 20.0% on Terminal-Bench 4.0 and from 3.2% to 24.0% on FrontierCode 2.0.
Lab claim: raises ExploitBench from 11.3% to 34.8%.
Sources
LABIntroducing SWE-2 Flashcognition.ai/blog/swe-2-flash · Nov 4, 2026LABCognition launch thread on Xx.com/i/status/2118069614521241739 · Nov 4, 2026Tracking ends Feb 2, 2027 unless the API or weights become available.