The distributed inference network

Fabriq runs large models
across many machines.

Point an existing OpenAI, Claude, or Ollama client at one endpoint. Requests route to a node that already has the model loaded.

OpenAI-compatible

Chat · Claude · Ollama

Open weights

Llama · DeepSeek · Qwen · FLUX

Apple Silicon

GPUs in progress

Per-token

usage-based pricing

Open models it can run

Llama 3.2DeepSeek V3.1Qwen3-235BKimi K2 ThinkingGLM-4.7gpt-ossFLUX.1Llama 3.2DeepSeek V3.1Qwen3-235BKimi K2 ThinkingGLM-4.7gpt-ossFLUX.1

The idea

The hardware already exists.
Most of it sits idle.

Models sharded across nodes, placed by topology.

Anyone can run a node. Any open model.

Pipeline and tensor parallelism; RDMA over Thunderbolt.

Verifiable execution. Not built yet.

Run a node locally, or call the managed endpoint. The same OpenAI-compatible API either way.