Faster inference from self-improving systems
We build systems that discover better ways to run open models. Our inference API puts that work in your hands.
Featured research
- Growing languagesReusable languages and abstractions for agents
- DeepSeek V4.1 Flash on a single B300Custom quantization and inference optimization on a single B300
- Dense Matmuls are faster when the inputs are sparseHow sparsity, power limits and clocks affect dense matmuls
- Kernel swappingHow kernel choices affect model-level quality
- Geometric on NVIDIA SOL-ExecBenchTop four across all four tracks, alongside entries from Databricks, Cursor, Qwen, Recursive and Tencent’s Hyra
Work with us
Inference
Run open models through Geometric’s optimized deployments
Get API access: InferencePartnerships
Bring Geometric’s implementations to your inference platform or customer workloads
Partner with us: PartnershipsGitHub app
Approve an optimization plan and review a pull request with verified results
Explore the app: GitHub appModels
Download an open model optimized with our performance systems
Browse models: Models