News · SI
Cognition's SWE-2 Flash is a cheaper, faster coding model for Devin, with weights held back for now
Cognition has launched SWE-2 Flash, a 276B-parameter (12B active) coding model mid-trained on Thinking Machines' Inkling-Small. It runs in Devin Desktop and CLI at one-fifth the price and about 3x the speed of SWE-2, and Cognition says it beats every American open-weight model on its evals. Citing cyber risk, it will share weights only with vetted partners.
01 / 04
Built on Thinking Machines' base model
SWE-2 Flash starts from Inkling-Small, Thinking Machines' 276B-parameter mixture-of-experts model with 12B active parameters[1]. For the first time, Cognition added a mid-training stage before reinforcement learning: a few trillion tokens of scientific articles, blogs, code and other sources, run in three stages so the scarcest, highest-quality data came last[1]. Cognition says most of the RL gains came from better and more varied training tasks rather than recipe changes[1].
02 / 04
The numbers
That extra training moved the base model a long way. Cognition reports Terminal-Bench 4.0 rising from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%[1]. On its published table, SWE-2 Flash:
- scores 24.0 on FrontierCode 2.0, just ahead of the 2.8T-parameter SWE-2 (23.8) and behind Claude Haiku 5.5 (27.7)[1];
- reaches 89.9 on SWE-Bench Pro v2 Hard and 89.2 on Terminal-Bench 2.1[1];
- trails DeepSeek V4.1 Flash and GLM 5.3 Flash on Terminal-Bench 4.0 (20.0 versus 42.3 and 41.8) and on DeepSWE 1.1[1].
Cognition's claim is that it beats every American open-weight model on leading evals, including ones 3x its size, rather than every open model worldwide[1].
03 / 04
Weights, but not for everyone
Cognition deliberately trained SWE-2 Flash for defensive security work, lifting its ExploitBench score from 11.3% to 34.8%[1]. Because the same skills could help attackers, it is not releasing the weights publicly yet. Instead it is building a program to share them securely with trusted industry partners and research labs, aimed at enterprises and governments that want to host and fine-tune the model on their own infrastructure, with more details promised in a few weeks[1].
04 / 04
Why it matters
Small, fast models now handle much of what coding agents do day to day, such as automations, code review, feature work and docs, and Cognition is betting teams will route that work to a cheaper model[1]. The gated-weights plan is the more interesting signal: it is another lab treating cyber capability as the reason to stop short of a full open-weight release.
FAQ
Frequently asked questions.
01What is SWE-2 Flash?
SWE-2 Flash is Cognition's new small coding model. It starts from Inkling-Small, a 276B-parameter mixture-of-experts model with 12B active parameters from Thinking Machines, then adds a mid-training stage on a few trillion tokens followed by large-scale reinforcement learning.
02Where can I use it?
It is available now in Devin Desktop and the Devin CLI. Cognition says it costs one-fifth as much as SWE-2 and runs about 3x faster, roughly 200 versus 60 tokens per second in its demo.
03Can I download the weights?
Not publicly. Cognition says SWE-2 Flash's cyber skills (ExploitBench rose from 11.3% to 34.8%) could help attackers, so it is building a program to share the weights securely with trusted industry partners and research labs, with details promised in a few weeks.
04How good is it?
On Cognition's own table it scores 24.0 on FrontierCode 2.0 (versus 3.2 for the Inkling-Small base and 23.8 for the far larger SWE-2), 20.0 on Terminal-Bench 4.0 and 89.9 on SWE-Bench Pro v2 Hard. It trails DeepSeek V4.1 Flash and GLM 5.3 Flash on Terminal-Bench 4.0 and DeepSWE 1.1, so it is not the top open-weight model overall.
That’s the story
Building with what’s new?
Spectrum Web Co ships products, places talent, builds social narratives and builds custom business tools as monthly services.
Related stories