Darkbloom – Private Inference on Idle Macs

Eigen Labs has unveiled a distributed inference network called Darkbloom. Its core idea is to connect idle Apple Silicon Macs as inference resources, providing low-cost AI inference without going through centralized GPU/API paths.

Key claims

• It estimates that 100M+ Apple Silicon devices have shipped since 2020, and most sit idle for 18+ hours a day.
• It aims for costs up to 70% cheaper than centralized alternatives.
• The API is OpenAI-compatible, supporting chat, image generation, and speech-to-text.
• In the top introduction it says operators keep 95% of revenue, while the operator economics section states platform fee 0% / operators keep 100%.

Security model

• Requests are encrypted on the client before being sent.
• The coordinator only routes ciphertext and cannot read the content.
• Nodes use keys generated inside Apple's secure hardware along with an attestation chain.
• The inference process is locked down with a hardened runtime, said to block debugger attachment and memory inspection.
• Responses are signed by the specific machine, making them traceable.

Implementation

• It says you can use the existing OpenAI SDK as-is and just change the base_url to https://api.darkbloom.dev/v1.
• It touts streaming, function calling, and compatibility with existing SDKs.
• The model catalog presented includes Gemma 4 26B, Qwen3.5 27B/122B MoE, MiniMax M2.5 239B, Cohere Transcribe, and others.

Pricing and operations

• A per-token pricing comparison table presents a 50% reduction versus OpenRouter.
• Image generation is listed at $0.0015/image, and speech-to-text at $0.001/minute.
• Operators can install a CLI on their idle Mac to run it as a launchd service, and it notes that a menu bar app is under development.