> ## Documentation Index
> Fetch the complete documentation index at: https://docs.darkbloom.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Darkbloom: Decentralized Private AI Inference

> Darkbloom routes AI to idle Apple Silicon Macs. Get an OpenAI-compatible API with hardware-level privacy, or earn by sharing your Mac's compute.

Darkbloom is a decentralized inference network that connects AI consumers with Mac owners running idle Apple Silicon hardware. As a consumer, you get access to frontier-quality models through an OpenAI-compatible API — with end-to-end encryption ensuring your prompts stay private. As a provider, you earn by running inference on your Mac during idle hours.

<CardGroup cols={2}>
  <Card title="Call the API" icon="code" href="/quickstart-consumer">
    Make your first inference request in under 5 minutes using any OpenAI SDK.
  </Card>

  <Card title="Become a Provider" icon="server" href="/quickstart-provider">
    Install the provider agent on your Apple Silicon Mac and start earning.
  </Card>

  <Card title="API Reference" icon="book" href="/reference/chat-completions">
    Full reference for all inference, account, and network endpoints.
  </Card>

  <Card title="Security & Trust" icon="shield-halved" href="/security/overview">
    Learn how Darkbloom ensures your prompts are never visible to providers.
  </Card>
</CardGroup>

## How it works

Darkbloom runs on a three-layer architecture: your requests go to a coordinator running in a Confidential VM, which routes them to attested provider machines running hardened inference processes on Apple Silicon.

```
Consumer (SDK / Web UI / curl)
    │
    │  HTTPS, OpenAI-compatible API
    ▼
Coordinator (Confidential VM, encrypted memory)
    │
    │  WebSocket — request re-encrypted to provider's key
    ▼
Provider (hardened process on Apple Silicon)
    │
    │  inference engine
    ▼
Apple Silicon GPU
```

The coordinator encrypts each request with the destination provider's public key before forwarding it. Only the hardened provider process — running with debugger attachment blocked and memory-reading APIs disabled — can decrypt and process the request.

## Available models

| Model                         | Size                 | Min RAM | Best for                   |
| ----------------------------- | -------------------- | ------- | -------------------------- |
| Gemma 4 26B 8-bit             | 26B MoE, 4B active   | 36 GB   | Fast multimodal requests   |
| Qwen3.5 27B Claude Opus 8-bit | 27B dense            | 36 GB   | Frontier-quality reasoning |
| Trinity Mini 8-bit            | 27B Adaptive MoE     | 48 GB   | Fast agentic inference     |
| Qwen3.5 122B MoE 8-bit        | 122B MoE, 10B active | 128 GB  | Best quality reasoning     |
| MiniMax M2.5 8-bit            | 239B MoE, 11B active | 256 GB  | SOTA coding, \~100 tok/s   |

<Note>
  Darkbloom is an experimental research prototype under active development. Expect rough edges and breaking changes. Do not use in production.
</Note>