Cognition Launches SWE-1.7 AI Model, Claims Near-Frontier Performance at Lower Cost

Cognition has launched SWE-1.7, a proprietary AI model built on a Kimi K2.7 base that delivers near-frontier coding performance at a fraction of the cost. Scoring 42.3% on the FrontierCode 1.1 benchmark, SWE-1.7 operates at $1.97 per task. To train the model, Cognition utilized a distributed multi-cluster setup spanning three continents, implementing compressed weight deltas and advanced entropy preservation techniques to stabilize long-horizon reinforcement learning.

SWE-1.7 model launch

• ▪SWE-1.7 scores 42.3% on Cognition's FrontierCode 1.1 benchmark, placing it behind GPT-5.5 at 43.0% and Claude Opus 4.8 at 46.5%.
• ▪SWE-1.7 costs $1.97 per task on the FrontierCode Main set, positioning it on a highly efficient spot on the cost-performance Pareto curve.
• ▪SWE-1.7 is available in Devin's Web, Desktop, and CLI clients via Cerebras hardware at 1000 tokens per second.
• ▪Cognition launched SWE-1.7, an AI model optimized for long-horizon asynchronous software engineering tasks, on July 8, 2026.
• ▪SWE-1.7 is built on a Kimi K2.7 base model and trained further using reinforcement learning.

Entropy preservation techniques

• ▪Cognition utilized the Muon optimizer and eliminated non-deterministic operations in the trainer to improve training stability.
• ▪Cognition developed sampling distribution replay to record tokens available at rollout time and renormalize probabilities in the trainer, bounding training-inference divergence.

Training stability improvements

• ▪Cognition addressed training stability by correcting the KL divergence mismatch between inference and training caused by asynchronous reinforcement learning.
• ▪Cognition trained SWE-1.7 across four datacenters on three continents, combining its own GPU clusters with rented capacity from Fireworks.

Multi-cluster distributed training

• ▪Cross-continental weight updates for a trillion-parameter model complete in 1 to 2 minutes, pausing inference for only 3 to 4 seconds.
• ▪To synchronize weights, the trainer uploads compressed weight deltas to cloud object storage, reducing transfer sizes by over 99%.
• ▪Inference-side failures are managed using NVIDIA Dynamo, which reroutes trajectories to healthy replicas without losing the full session state.

Fault tolerance mechanisms

• ▪Cognition built an automated data-quality pipeline that runs execution tests, filters low-signal tasks, and hardens tasks against reward-hacking.
• ▪The trainer checkpoints asynchronously to local disk every step and replicates shards to peer nodes, allowing rapid state recovery during hardware failures.

Data curation pipeline

• ▪Cognition curated training data by focusing on tasks where the model solves only a low fraction of the time to maximize learning signal.

3 sources