Cognition drops SWE-1.7, its bargain coding powerhouse

For months, the standard playbook for AI coding agents was simple: hook your product up to the most expensive closed-source frontier model and pass the eye-watering API bill to the user.

Cognition is officially rewriting that script. On July 8, the team behind Devin launched SWE-1.7, its new in-house model built on top of China's open-source Kimi K2.7 base. The model delivers frontier-level performance comparable to GPT-5.5 and Claude Opus at a fraction of the operating cost. It is available right now across Devin’s Web, Desktop, and CLI interfaces, running on Cerebras hardware at blistering speeds of 1,000 tokens per second.

On FrontierCode 1.1, SWE-1.7 leaped from its base model's 30.1% to 42.3%, essentially matching GPT-5.5's 43.0%. It also posted 81.5% on Terminal-Bench 2.1 and 77.8% on SWE-Bench Multilingual, completely leaving its predecessor SWE-1.6 in the dust.

Squeezing extra juice from post-trained models

Conventional industry wisdom suggested that once an open-source model has undergone heavy reinforcement learning post-training, you quickly hit a 'post-training ceiling' where further RL yields diminishing returns.

Cognition proved that ceiling is fake. To keep training stable over massive runs, they tackled entropy collapse and inference-trainer drift head-on. By logging kept-sets during top-p rollouts and replaying those sampling distributions in the trainer, they kept entropy constant while filtering out unhelpful gradient noise. Swapping in the Muon optimizer and stripping out non-deterministic operations kept the scale predictable.

For complex engineering tasks that easily blow past raw context limits, the model also learned self-compaction. It summarizes its active working state, resumes from the summary, and uses alternating length penalties to stay concise without dropping accuracy.

Scrappy multi-cluster engineering across three continents

Instead of waiting in line for a single massive cluster of 100,000 chips, Cognition took advantage of how RL naturally splits apart: only the trainer needs ultra-high bandwidth, while inference rollout engines can live anywhere.

They wired up four datacenters across three continents, stitching together their own hardware with third-party providers like Fireworks. To prevent cross-continental weight updates from bottlenecking the run, they transmitted compressed weight deltas that shrank transfer payloads by over 99%.

Using cloud object storage as the single source of truth, 1T-parameter model weights sync across continents in 1 to 2 minutes, pausing inference for merely 3 to 4 seconds while worker engines swap weights in-place. If hardware fails, NVIDIA Dynamo instantly reroutes agent sandboxes so partial trajectories are never lost.

Why it matters

For developers using Devin, the immediate payoff is sheer execution speed and better task endurance. Hitting 1,000 TPS via Cerebras while keeping reasoning tight means long-horizon bug fixing and refactoring should feel remarkably snappier, making it well worth spinning up for a test run today.

For the broader AI ecosystem, this is a masterclass in capital efficiency. Cognition demonstrated that specialized agent startups do not need to spend hundreds of millions pre-training baseline models from scratch. Taking a world-class open-source foundation like Kimi K2.7 and investing heavily in bespoke RL plumbing and distributed infrastructure can deliver frontier capabilities at startup scale.

We think the cleverness lies in treating compute as a distributed commodity rather than a monolithic cage. The question mark, however, is long-term leverage: heavy specialization on another team's base architecture means when foundation architectures shift, you have to continually re-engineer your downstream pipeline to stay ahead.