Today we’re releasing SWE-1.5, the latest in our family of models optimized for software engineering. It is a frontier-size model with hundreds of billions of parameters that achieves near-SOTA coding performance. It also sets a new standard for speed: we partnered with Cerebras to serve it at up to 950 tok/s – 6x faster than Haiku 4.5 and 13x faster than Sonnet 4.5. It is now available in Windsurf.

Read the full research announcement on the Cognition website.

Why Speed Matters

When you’re coding, waiting 20 seconds for your AI to respond breaks your flow. We built SWE-1.5 to eliminate that frustration. It delivers frontier-level coding intelligence at speeds up to 13x faster than leading models, so you can stay in the zone while tackling complex tasks.

Built for Real Development Workflows

SWE-1.5 is optimized specifically for Windsurf’s agent experience. We trained it end-to-end on real coding tasks, continuously dogfooding it internally to ensure it handles the messy, complex scenarios you encounter daily. This tight integration means the model understands how to use Windsurf’s tools effectively and responds in ways that feel natural to your workflow.

When a model runs 10x faster, everything else becomes a bottleneck. We rewrote critical components like lint checking and command execution to keep up, reducing overhead by up to 2 seconds per step. These improvements benefit all models in Windsurf, not just SWE-1.5.

Trained on Real-World Tasks

SWE-1.5 was trained on diverse, real-world coding scenarios that mirror what you actually do in Windsurf. We focused on teaching the model to write clean, maintainable code, not just code that passes tests. This means less verbose output, fewer unnecessary try-catch blocks, and solutions that follow best practices.

We worked with senior engineers and open-source maintainers to ensure the model learns from high-quality examples across many languages and frameworks. The result is an AI that understands not just what to code, but how to code well.

Powered by Cerebras

To deliver unprecedented speed, we partnered with Cerebras, achieving inference speeds up to 950 tokens per second. We also built a custom request priority system to ensure smooth, responsive agent sessions even under load. The result is an experience that feels instant. Tasks that used to take minutes now complete in seconds.

Performance That Matters

On SWE-Bench Pro, a challenging benchmark of real-world coding tasks, SWE-1.5 achieves near-frontier performance while completing tasks in a fraction of the time.

More importantly, SWE-1.5 is now the daily driver for many of our engineers. Here’s what they’re using it for:

• Exploring large codebases: Quickly understand unfamiliar code (SWE-1.5 powers our new Codemaps feature)
• Full-stack development: Build complete features from frontend to backend
• Infrastructure work: Edit Kubernetes manifests, Terraform configs, and other complex YAML files without memorizing field names

Tasks that used to take 20+ seconds now complete in under 5 seconds, ultimately keeping you in flow.

Try It Today

SWE-1.5 proves you don’t have to choose between speed and intelligence. It’s available now in Windsurf! Just select it from the model picker and experience the difference yourself.

Whether you’re exploring a new codebase, building a full-stack feature, or tackling infrastructure work, SWE-1.5 keeps pace with your thinking. This is just the beginning, we’re continuing to push the boundaries of what’s possible in AI-assisted coding.