An Eigen Labs project · Research preview

Private inference on verified Apple Silicon.

Darkbloom routes AI requests to hardware-attested Macs. Your prompts are end-to-end encrypted — the machine running your model can never read them.

Start using Darkbloom → Read the docs

E2E

X25519 encrypted

4-layer

Attestation

Platform fee

[OI]

Compatible API

01 — Why Darkbloom

The provider can’t see your prompts. Provably.

The core problem with running AI on someone else’s Mac: the owner has root and physical custody. Darkbloom eliminates every software path through which inference data could be observed — then proves it, cryptographically.

/ 01

Encrypted before it leaves

Every request is encrypted with the provider’s X25519 public key. The coordinator routes ciphertext — it never sees plaintext either.

/ 02

Verified hardware

Secure Enclave signatures, MDM cross-checks, and Apple Managed Device Attestation verify each Mac’s security posture — publicly, at /v1/providers/attestation.

/ 03

Hardened at the OS level

Debuggers denied at the kernel (PT_DENY_ATTACH), memory reads blocked by Hardened Runtime, memory isolation via Hypervisor.framework.

02 — How it works

Three moving parts, zero trust between them.

YOU

Consumer

SDK, web console, or curl. Any [OI]-compatible client — just change the base URL.

HTTPS

ROUTING

Coordinator

Auth, billing, and encrypted routing to a verified provider. Sees ciphertext only.

E2E ENC

COMPUTE

Provider Mac

A hardened Rust process decrypts and runs inference on Apple Silicon GPU. Nothing leaks out.

03 — Models & pricing

Curated catalog. Per-token, in USD.

Models are selected from a verified catalog — the coordinator only routes to models it has checked. 0% platform fee during research preview; providers keep 100%.

Model	Type	Size	Input / 1M	Output / 1M
Gemma 4 26B	Text · MoE	26B (4B active)	$0.065	$0.20
Qwen3.5 27B	Text · Dense	27B	$0.10	$0.78
Qwen3.5 122B	Text · MoE	122B (10B active)	$0.13	$1.04
MiniMax M2.5	Text · MoE	239B (11B active)	$0.06	$0.50
FLUX.2 Klein 4B	Image	4B diffusion	$0.0015 / image
Cohere Transcribe	Speech-to-text	2B conformer	$0.001 / audio-min

04 — The API

Drop-in compatible.

Works with any [OI] SDK by changing the base URL. The Anthropic Messages API is also supported at /v1/messages.

✓Chat, completions, transcription, and image endpoints
✓Streaming with trust metadata per response
✓Public, verifiable attestation feed

Full API reference →

curl

# Any [OI]-compatible client — just point it here:
curl https://api.darkbloom.dev/v1/chat/completions \
  -H "Authorization: Bearer $DARKBLOOM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemma-4-26b-8bit",
    "messages": [
      {"role":"user","content":"Hello"}
    ],
    "stream": true
  }'

05 — For Mac owners

Your idle Mac is the data center.

Over 100 million Apple Silicon Macs sit idle most of each day — 64–512 GB of unified memory, up to 819 GB/s of bandwidth. Darkbloom connects that capacity directly to demand.

✓Zero-prerequisite installer: binary, Python 3.12, vllm-mlx, and Secure Enclave tooling bundled
✓Native SwiftUI menu bar app with scheduling and earnings
✓Providers connect outbound — no port forwarding, no static IP

Become a provider →

/ ECONOMICS

Keep 100% of inference revenue

0% platform fee during research preview. Electricity on Apple Silicon runs $0.01–0.03/hour depending on workload — the rest is yours.

/ SCHEDULING

Serve only when idle

Time-based availability windows shut the backend down outside your hours, freeing GPU memory. Configured in the app or provider.toml.