Two ways to use Malibu
Build something.
Or help run it.
Developers bring the requests. Mac owners bring the compute. Choose your side of the network.
The OpenAI SDK is the Malibu SDK.
Point any OpenAI-compatible client at api.malibu.tech/v1. Streaming, tool calling, structured output, sticky conversations, and signed inference receipts — all live today.
client = OpenAI(
base_url="https://api.malibu.tech/v1",
api_key="mp_…"
)
Cheaper by design
Distributed Apple Silicon undercuts hyperscaler GPU pricing on open models. Pay per token, not per H100-hour.
Cooperative trust
Prompts route to individual Macs as plaintext — not confidential compute. Signed receipts bind the response. See the security docs.
Signed receipts
Every response carries an Ed25519 receipt you can verify offline with malibu-verify.
Open model catalog
Llama, Qwen, GPT-OSS, and more — one endpoint, no per-provider onboarding.
Drop-in OpenAI SDK
Change base_url, keep your code. Streaming, tool calling, structured output all work.
Prepaid credits
Top up once, no subscription. GitHub OAuth for the API key; no wallet required to start.
Your Mac already has the hardware.
Download Malibu for macOS, click Launch Provider, and your Apple Silicon Mac joins the pool. Malibu picks the right open model for your memory tier, serves when you're idle, and credits you 90% of billed inference. USDC payout ships at public beta. Or point a coding agent at get.malibu.tech/skill.md. $MALIBU emission is planned at mining launch.
Credits accrue now
Every billed prompt credits 90% to the provider. USDC-on-Base payout ships at public beta. $MALIBU minting is not live yet.
Runs on pennies
Apple Silicon draws 20–40W under inference load. Passive income, not a power bill.
Your Mac stays yours
Cooperative scheduling — serving interrupts for milliseconds. No datacenter, no rack.
Zero configuration
Malibu picks the right open model for your memory tier at install time. No tuning. Invite required to register.
Early providers at launch
When mining ships, the first 90 days are designed to mint 5× more $MALIBU per prompt. Not paying today.
Pause anytime
Toggle serving on when idle, off when you're working. No commitment, no schedule.
Frontier open models, MLX-native.
A curated catalog of the strongest open-weight models — quantized for Apple Silicon, matched to provider hardware at install time. Qwen2.5 7B, Qwen3.5 35B, Llama 3.3 70B, and gpt-oss 120B are in the catalog and serve when a matching-RAM Mac is online. Pricing tracks the OpenRouter frontier — compare the live rate card, no subscription, no seat license.
A single Mac Studio Ultra already runs frontier open models — gpt-oss 120B — from unified memory, no datacenter. Pooling Macs to serve models beyond any single machine is next on the roadmap; a verifiable receipt is worth the most exactly at that scale. Not live yet.
The network is the token
Today, billed inference credits providers 90%. USDC payout ships at public beta. At token launch, every dollar of buyer spend is designed to split: providers get paid, supply burns, reserves grow.
Hold $MALIBU, own the network — that's the launch design. Live billing is 90/10 credits, not the burn-and-reserve split.
Three things happen the moment a prompt clears under token economics. Not the live 90/10 USDC rail.
The Mac gets paid.
70¢ of every dollar goes to the person whose Mac ran the prompt, in USDC.
Supply shrinks.
18¢ buys $MALIBU on the open market and burns it. Every prompt leaves fewer tokens in circulation.
The floor rises.
12¢ goes into a reserve that backs every new $MALIBU. Macs keeping the network up earn the new tokens.