# Geodd

> AI inference platform that serves models through one unified API, continuously improved by hardware-specific AI agents that develop and deploy optimized kernels for NVIDIA, AMD, and Tenstorrent accelerators.

Geodd is an AI inference company that serves models through a single unified API while continuously improving performance using hardware-specific AI agents. The platform combines model serving with accelerator kernel development, using production traffic to guide ongoing optimizations across NVIDIA, AMD, and Tenstorrent hardware. It is GDPR-ready, SOC 2 pending, and operates active regions in the US East and EU Norway, with APAC expansion underway.

## What It Is

Geodd provides serverless inference and dedicated GPU deployment for AI teams that need consistent, production-grade model execution. Rather than treating inference as a static service, Geodd runs a continuous improvement loop: AI agents observe real workloads, identify bottlenecks, generate hardware-specific kernel code, test it in staging, and deploy verified improvements automatically into serving. The result is a platform where performance improves over time based on the actual conditions of production traffic.

## How the Kernel Development Loop Works

The core differentiator is a five-stage cycle that connects serving observations to kernel improvements:

- **Observe real workloads** — execution graphs, context lengths, and batch sizes reveal where inference spends time.
- **Write better kernels** — hardware-specific LLMs (Mosaic for NVIDIA/CUDA, Druze for AMD/ROCm, Strata for Tenstorrent/TT-Metalium) generate and refine code targeting identified bottlenecks.
- **Test against the workload** — correctness and performance are verified under the conditions that exposed the problem.
- **Deploy and measure** — verified improvements are released into serving and monitored.
- **Feed the next cycle** — successful kernels and measured results inform the next round of experiments and further development of the specialist models.

Druze and Strata are listed as coming soon; Mosaic (NVIDIA/CUDA) is the active specialist model with published benchmark results.

## Platform and Deployment Options

Geodd offers two primary inference modes:

- **Serverless inference** — token-based, usage-priced API access to a catalog of models including DeepSeek V4 Flash, GLM 5.2, GPT-OSS-120B, Kimi K2.6, Gemma 4 31B, ByteDance Seed/Seedream/Seedance families, and others.
- **Dedicated deployment** — isolated infrastructure for teams that need reserved capacity, custom SLAs, or multi-region clusters.

The platform also includes **Deploypad**, described as instant model orchestration, and an **Optimized Model Engine** for high-performance execution. Geodd publishes free local inference runtimes with model-specific optimized kernels for teams that want to run on their own hardware; these are separate from the internal specialist LLM weights.

## Developer Integration

Geodd is fully compatible with the OpenAI SDK. Switching providers requires only changing the `base_url` to `https://api.geodd.io/inference/v1` and supplying a Geodd API key — no migration of existing code. The platform provides real-time token usage and observability, and operates a Zero Data Retention (ZDR) policy for standard API prompts and outputs. A unified API covers both serverless inference and dedicated GPU compute.

## Update: ByteDance Models and EU Region Launch

Recent blog posts and the homepage banner confirm several active product developments. ByteDance Seed, Seedream, and Seedance model families are now live on Geodd, covering language/agent workloads, image generation, and AI video generation. EU serverless inference launched on GPU infrastructure hosted in Norway. Geodd has also announced a partnership with Opper AI to make its inference infrastructure available through the Opper AI gateway. The changelog covers product, API, infrastructure, billing, and security updates from April through September 2026, indicating active development cadence.

## Features
- Serverless inference API
- Dedicated GPU deployment
- OpenAI SDK compatibility
- Hardware-specific AI kernel development agents
- Mosaic specialist model for NVIDIA/CUDA optimization
- Druze specialist model for AMD/ROCm (coming soon)
- Strata specialist model for Tenstorrent/TT-Metalium (coming soon)
- Deploypad model orchestration
- Optimized model engine
- Free local inference runtimes with optimized kernels
- Multi-regional deployment (US East, EU Norway, APAC expansion)
- Zero Data Retention (ZDR) policy
- GDPR-ready data handling
- SOC 2 pending
- Real-time token usage and observability
- Unified API for serverless and dedicated compute
- Rate limit management
- Startup credits program

## Integrations
OpenAI SDK, Opper AI, NVIDIA CUDA, AMD ROCm, Tenstorrent TT-Metalium, DeepSeek, ByteDance Seed/Seedream/Seedance, GLM, Gemma, Kimi

## Platforms
WEB, API, CLI

## Pricing
Freemium — Free tier available with paid upgrades

## Links
- Website: https://geodd.io/
- Documentation: https://geodd.io/docs
- EveryDev.ai: https://www.everydev.ai/tools/geodd