Cognition Releases SWE-2 Flash: 5x Cheaper, and 20% of Terminal Tasks Solved Versus 0.9%

Most of the day-to-day work of coding agents doesn't need the most expensive model. Reviewing a pull request, updating a README after a config change, closing a small ticket in the tracker: a flagship is overkill for all of these, and its price and slowness add up fast. Cognition has released SWE-2 Flash, which it calls a fully American coding model.

It is five times cheaper than the flagship SWE-2 and three times faster: about 200 tokens per second versus 60 (a token is a chunk of text, roughly a syllable or two). Quality has improved too. On Terminal-Bench 4.

0, where the agent runs commands in a terminal on its own and solves practical tasks such as analyzing biological data, the model completes 20% of the assignments. The base Inkling-Small managed 0.9%.

What you get in speed and price

SWE-2 Flash is already available in Devin Desktop and Devin CLI. The speed difference shows in long agent sessions, where the model reads files, runs tests and edits code dozens of times in a row. Every pause becomes waiting time, and a response that arrives three times faster cuts that wait noticeably.

The base is Inkling-Small from Thinking Machines: 276 billion parameters in total, of which 12 billion are active at a time. For comparison, Beam 1.0 has 501 billion, Muse Spark Open has 816 billion, and SWE-2 has about 2.8 trillion. According to Cognition, Flash beats all American open-weight models on leading benchmarks, including those three times its size.

Where the gain came from

The main new ingredient in the recipe is mid-training before reinforcement learning. Cognition had not used this stage before. The team assembled a dataset of several trillion tokens: scientific papers, blogs, code and other sources. Training ran in three stages, with the rarest and highest-quality data shown to the model at the end.

To choose the data proportions, the engineers compared different mixes. They measured how many tokens each needed to reach the same error on a held-out validation set as the baseline run. They built that set themselves, and it reflects real development tasks.

Large-scale reinforcement learning followed. The recipe itself changed little, and according to Cognition, most of the gain came from the tasks. The team checked the instructions, environment setup and verification scripts, and also added new types of tasks so the model would learn to reason more broadly.

Results compared with the base Inkling-Small:

• Terminal-Bench 4.0: from 0.9% to 20.0%;
• FrontierCode 2.0: from 3.2% to 24.0%;
• DeepSWE 1.1: from 17.4% to 60.1%;
• ExploitBench: from 11.3% to 34.8%.

On FrontierCode 2.0, Flash scored 24.0% versus 23.8% for the flagship SWE-2. On Terminal-Bench 4.0 the flagship is still ahead (27.3% versus 20.0%), and on DeepSWE 1.1 the gap is larger (73.0% versus 60.1%).

Cyber risks and closed weights

Cognition trained the model separately on cybersecurity so that it could help defenders. Its result on ExploitBench, a test of exploit creation, roughly tripled. But the same skills are useful to attackers, and the company cites the recent surge in cyberattacks as the reason not to publish the weights openly.

Instead, it is preparing a program under which vetted industry partners and research labs will receive the weights. They will be able to fine-tune the model and run it in closed environments, including on their own servers. Details are promised in a few weeks.

What to do with this

If your agents mostly review code, write documentation and close small tasks, it makes sense to run them through Flash and compare the results with an expensive model. Until November 30 you can do this for free. Teams with strict privacy requirements should watch the terms of the weight-access program.

The caveats are obvious. All the figures come from Cognition itself, and there are no independent verifications yet. The flagship SWE-2 remains stronger on hard tasks, and a model that solves a fifth of terminal tasks will not replace an engineer anytime soon. Still, the combination of price, speed and quality makes it a serious candidate for high-volume automated work.