Cognition launches SWE-1.5 with 950 tokens per second

Cognition has launched SWE-1.5, a new large language model designed for software engineering tasks. With hundreds of billions of parameters, SWE-1.5 is positioned as a top-tier coding model, achieving near state-of-the-art performance on coding benchmarks. The model is publicly available for users through the Windsurf platform, targeting developers and engineering teams seeking higher speed and accuracy in code generation and analysis.

The release is notable for its partnership with Cerebras, enabling inference speeds up to 950 tokens per second, which is six times faster than Haiku 4.5 and thirteen times faster than Sonnet 4.5.

SWE-1.5 brings several technical advancements. It has been trained using end-to-end reinforcement learning in realistic coding environments, leveraging a custom Cascade agent harness and a robust infrastructure powered by thousands of GB200 NVL72 chips. This model is one of the first public production deployments on the new GB200 hardware generation, allowing for high concurrency and rapid processing. The training included rigorous internal testing, continuous iteration on tools and prompts, and a focus on real-world code quality through a combination of classical testing, rubrics, and agentic grading. Early users within the company report substantial improvements in workflow speed, with common tasks like Kubernetes manifest editing now performed in under five seconds.

Cognition, the company behind SWE-1.5, has a background in developing agentic AI systems for software engineering. Their approach integrates model development, infrastructure, and product experience, with a focus on optimizing both speed and coding quality. With this release, Cognition continues its strategy of tightly coupling AI model research with practical engineering tools, aiming to set new standards in the automated software development industry.

Source