SWE-2 Flash

Overview

SWE-2 Flash is a coding model released by Cognition on November 4, 2026, built as a 276B-parameter mixture-of-experts model with 12B active parameters on top of Thinking Machines' Inkling-Small. Cognition reports about 3x the speed of its larger SWE-2 model (roughly 200 tokens/s versus 60 tokens/s in a simulated replay) at about one-fifth the price. It runs in Devin Desktop and the Devin CLI, and Cognition says it is not sharing the weights publicly yet, planning a program for trusted industry partners and research labs.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

• 276B total / 12B active parameter MoE, post-trained from Inkling-Small
• Vendor-reported scores: Terminal-Bench 2.1 89.2, SWE-Bench Pro v2 Hard 89.9, DeepSWE 1.1 60.1
• About 3x faster than SWE-2 (about 200 vs 60 tokens/s in a simulated replay)
• Priced at about one-fifth of SWE-2
• Available in Devin Desktop and Devin CLI; weights not publicly released

Use cases

• Low-latency coding-agent turns inside Devin Desktop
• Terminal and repository tasks driven from the Devin CLI
• Cheaper sub-agent runs alongside the larger SWE-2 model

Pricing

Free for paid Devin users through November 30, 2026; afterwards about one-fifth the price of SWE-2 (no absolute price published).

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released

4 Nov 2026

Record updated

6 Nov 2026

Record last checked

6 Nov 2026

• cognition.com/blog/swe-2-flash
• cognition.com/blog