Geometric provides an inference API for open-weight AI models, directly and through partners such as OpenRouter. Its security page states that European inference deployments run on Nebius and Verda across the UK, France and Finland. European traffic is never routed to US capacity, and customers can request that EU traffic remain on EU servers.
Geometric says prompts and outputs are never written to disk or logs or used for model training or evaluation. Its inference service uses short-lived in-memory caching. Traffic between its gateway and GPU servers travels through encrypted WireGuard tunnels.
Privacy Policy — GeometricListed 2 Nov 2026 · Geometric · Web pageGeometric’s privacy policy names Nebius and Verda as providers of its GPU infrastructure in the UK, France and Finland.
Security — Geometric
Listed 2 Nov 2026 · Geometric · Web page
Geometric states that its European inference deployments run on Nebius and Verda in the UK, France and Finland, with European traffic never routed to US capacity.
Security — GeometricListed 2 Nov 2026 · Geometric · Web pageGeometric states that its European inference deployments run on Nebius and Verda in the UK, France and Finland, with European traffic never routed to US capacity.
Geometric develops systems that optimize how open-weight AI models run and offers those deployments through an inference API for businesses and developers. Its work covers GPU kernels, inference engines, quantization and self-improving optimization methods. The API supports OpenAI and Anthropic client libraries. Geometric also distributes optimized models, works with inference platforms on customer workloads, and offers a GitHub app for reviewing proposed optimizations and their results. Its service is operated by Vision Inference Ireland Limited, within the Vision Inference, Inc. group.
Geometric announces single-B300 deployment of DeepSeek V4.1 Flash
Geometric announced a custom quantization algorithm and inference stack that run DeepSeek V4.1 Flash on one NVIDIA B300, with access available by request. Its internal testing measured 260 output tokens per second with a 10,000-token input and one request at a time.
Geometric provides an inference API for open-weight AI models and develops systems to optimize their execution. · Infrastructure
How they use Nebius AI Cloud
Geometric provides an inference API for open-weight AI models, directly and through partners such as OpenRouter. Its security page states that European inference deployments run on Nebius and Verda across the UK, France and Finland. European traffic is never routed to US capacity, and customers can request that EU traffic remain on EU servers.
Geometric says prompts and outputs are never written to disk or logs or used for model training or evaluation. Its inference service uses short-lived in-memory caching. Traffic between its gateway and GPU servers travels through encrypted WireGuard tunnels.
Privacy Policy — GeometricListed 2 Nov 2026 · Geometric · Web pageGeometric’s privacy policy names Nebius and Verda as providers of its GPU infrastructure in the UK, France and Finland.
Security — Geometric
Listed 2 Nov 2026 · Geometric · Web page
Geometric states that its European inference deployments run on Nebius and Verda in the UK, France and Finland, with European traffic never routed to US capacity.
Security — GeometricListed 2 Nov 2026 · Geometric · Web pageGeometric states that its European inference deployments run on Nebius and Verda in the UK, France and Finland, with European traffic never routed to US capacity.
Geometric develops systems that optimize how open-weight AI models run and offers those deployments through an inference API for businesses and developers. Its work covers GPU kernels, inference engines, quantization and self-improving optimization methods. The API supports OpenAI and Anthropic client libraries. Geometric also distributes optimized models, works with inference platforms on customer workloads, and offers a GitHub app for reviewing proposed optimizations and their results. Its service is operated by Vision Inference Ireland Limited, within the Vision Inference, Inc. group.
Geometric announces single-B300 deployment of DeepSeek V4.1 Flash
Geometric announced a custom quantization algorithm and inference stack that run DeepSeek V4.1 Flash on one NVIDIA B300, with access available by request. Its internal testing measured 260 output tokens per second with a 10,000-token input and one request at a time.