Black Forest Labs' image generation and editing model with bounding-box layouts, multiple reference images and 4K output
SWE-2
New· 4Cognition's software-engineering model for Devin, post-trained from the open-weights Kimi K3
About SWE-2
A coding model that counts what each attempt costs, as well as whether it works, is how Cognition describes SWE-2, released on 10 September 2026. It is Cognition's software-engineering model and powers the Devin tools: Devin Desktop and the Devin CLI carry it from day one, and Devin Web and Fusion come afterwards. Access is through Devin, not through a separate API announcement.
How it was trained
SWE-2 starts from Kimi K3, the open-weights model from Moonshot AI, which Cognition post-trained with reinforcement learning using the recipe of its earlier SWE-1.7. The training objective rewards a successful result and also weighs what each attempt cost, in compute spent and time taken. The model has three reasoning-effort levels, medium, high and max, which Cognition trained together in a single run.
What Cognition reports
All results below come from Cognition and are taken at whichever effort setting scored highest. On FrontierCode 1.1 Main, the company reports 50.0 percent, within a point of one leading rival model and, it says, at a much lower cost. It reports 73.0 percent on DeepSWE 1.1 and 92.8 percent on Terminal-Bench 2.1. On the newer Terminal-Bench 4, SWE-2 reaches 27.3 percent, well behind two rival models that score above 55 percent. The cost comparison assumes list prices with public discounts.
What is not stated
The announcement gives no context window, price or licence, so none can be reported here.
Who it is for
SWE-2 serves people who already work in Devin and want a model tuned to the product, whose efficiency goal is stated openly. Its mixed benchmark picture, strong on some coding tests and weaker on terminal-heavy ones, is useful to keep in mind when deciding which tasks to hand it.
TrustList summary
SWE-2 is a software-engineering model from Cognition, released on 10 September 2026. It is available in Devin Desktop and the Devin CLI from launch, with Devin Web and Fusion to follow. It is offered inside Devin and was not announced as a standalone API.
Main features
- Coding agent model: built to power Cognition's Devin tools for software tasks.
- Reinforcement-learning training: post-trained from Moonshot AI's open-weights Kimi K3, extending the training approach used for the earlier SWE-1.7.
- Cost-aware objective: training rewards success while also counting inference spend and time per attempt.
- Three effort levels: medium, high and max, trained together in one run.
Key facts
- Released: 10 September 2026
- Maker: Cognition
- Base model: Kimi K3 from Moonshot AI
- Reported results: Cognition cites figures on FrontierCode, DeepSWE and Terminal-Bench, including a lower result on Terminal-Bench 4 than two rival models; all are vendor figures taken at whichever effort setting scored highest
- Not stated: context window and licence
Summary written by TrustList from the business's own published information, Nov 2026.
Benchmarks & AI stats
| Benchmark | Official | Community avg |
|---|---|---|
| DeepSWEcode | 73 % | - |
| FrontierCode 1.1 Maincode | 50 % | - |
| Terminal Bench 2.1agent | 92.8 % | - |
| Terminal-Bench 4.0agent | 27.3 % | - |
“Official” values are editor-approved and feed the ranking. “Community avg” is the mean of member submissions (shown for transparency; it never affects the ranking until an editor approves a value).
Sign in to add a benchmark score for this model.
SWE-2 launches
Trust Score
An earned signal from verification, reviews, awards, transparency and engagement - the vendor can't buy it.
Updated 9/16/2026
What the business tells TrustList
- Founded
- 2026 (0 years) evidence sent
Request a demo or quote from SWE-2
Reviews
Questions & answers
No questions yet. Be the first to ask about SWE-2.
SWE-2 alternatives
Similar listings buyers compare against this one.
JetBrains' open 12B mixture-of-experts coding model with 2.5B parameters active per token under Apache 2.0
Google's open multimodal embedding model that maps text, images, audio and video into one vector space, under Apache 2.0
Perplexity's open-weight decision model that picks among given options for routing and checks in AI systems
Microsoft AI's real-time speech-to-text model for voice agents, transcribing 60 languages with continuous language detection
Microsoft AI's text-to-speech model for keeping one voice across 23 languages, with voice cloning from short audio
Aleph Alpha's open-weight English-German mixture-of-experts model for self-hosting, released under Apache 2.0 in October 2026
Open-source 2B decision model from AWS's Strands Agents project that picks options and returns scores with a reliability figure
More on TrustList
Everything here links back to the same verified catalogue. Pick your next stop.
- Write a reviewHelp other buyers decide
- CompaniesAgencies, consultancies and IT service providers, ranked by verified reviews.
- ProductsSoftware and SaaS with pricing, features, integrations and alternatives.
- ArticlesGuides, comparisons and research from the editorial desk and the community.
- AwardsAnnual recognition decided by verified reviews and an independent jury.
- LaunchesNew products and releases, voted up by the community every day.
- RequestsBuyers describe what they need; vendors respond directly.
- PeopleReviewers, authors and makers with public profiles.
- ComparePut up to four listings side by side before you shortlist.