​

Malibu Litepaper

The world’s biggest AI cloud already exists. Malibu turns it on.

​

The premise

Roughly two hundred million Apple Silicon Macs have shipped since 2020 (a directional industry estimate — methodology). Every one of them has unified memory that can hold open-weight models, and most of them sit idle for much of the day. Fleet idle-time figures are planning assumptions, not audited telemetry — see Benchmarks & methodology. Malibu turns Apple Silicon Macs into a coordinated inference network. Admitted providers accrue credits for the verified inference they serve. Token-protocol conversion of buyer dollars into $MALIBU demand is planned — see Economics. Malibu doesn’t build new AI infrastructure. It activates the infrastructure already sitting on millions of desks.

​

The opportunity

AI inference is a large and fast-growing market, and a handful of companies own most of it. Open-source models — Llama, Qwen, gpt-oss — match closed-model quality on many chat, coding, and retrieval workloads (not every reasoning or multimodal eval). Supply hasn’t. Buyers often pay a large markup over the underlying compute. New datacenter capacity takes billions of dollars and years to permit. Meanwhile hundreds of millions of AI-capable Macs sit on desks running at low utilization (the “four percent” figure is a planning assumption, not measured fleet telemetry). Every prior decentralized AI network — Bittensor, Akash, Gensyn, Prime Intellect — hit the same wall: buyers won’t come without providers; providers won’t stay without buyers. A decade of permanent bootstrap. Malibu’s bet is to pay providers to be ready before buyers arrive (PoMW + emission at v0). Today’s live product is the inference marketplace (credits and receipts) on a small invite-gated pool. USDC-on-Base payout is planned for public beta and is off today. Check Network status for the ledger.

​

Why now

Five years ago this network wasn’t possible. Today, four preconditions are simultaneously true for the first time:

• Apple Silicon. Unified-memory Macs capable of serving large open-source models are shipping in consumer hardware.
• Open-source frontier models. Llama, Qwen, gpt-oss now match closed-model quality on many workloads. Verified inference on open weights is finally worth paying for.
• Cheap inference verification. TOPLOC (Prime Intellect, 2025) is designed to verify buyer responses at roughly one percent of serving cost — planned v1 on Malibu, not live. Receipts are live today.
• Stablecoin payments. USDC on Base can make global micropayments settlement-final in seconds, at sub-cent cost. Malibu’s USDC payout rail is built but not live yet.

Malibu is not early to any of these. It is the network that combines all four.

​

What we’ve shipped

Malibu is a live coordinated inference network today — distributed Apple Silicon providers, centralized routing and billing. Real coordinator, real providers, real inference requests, real credit ledger. Admission is private pre-beta: a valid invite is required to register a new provider. Public API buyers use a free daily token quota and are not charged money in pre-beta. USDC-on-Base payout is planned for public beta.

​

Inference marketplace — Live

• Coordinator service. Live Provider admission, real-time request routing, the credit ledger, and receipt-based settlement.
• Gateway. Live OpenAI-compatible /v1/chat/completions and /v1/models at api.malibu.tech/v1, plus /v1/usage (daily quota) and /v1/receipts/{request_id} (receipt status). Public, no-key /v1/status, /v1/rate-card, and /v1/stats/*. Beta Experimental OpenAI Responses (/v1/responses) and Anthropic Messages (/v1/messages) facades over chat completions.
• Provider onboarding. Beta The notarized Malibu app installs the provider, benchmarks the Mac, and picks the best paid model that fits. The malibu-cli command covers the same flow from a terminal, including headless installs over SSH. Invite required for a fresh register (malibu.tech/host). Already-running providers update without a new invite.
• Agent-readable onboarding. Live Point a coding agent at get.malibu.tech/skill.md to install, inspect, recover, update, or uninstall the provider CLI, and to connect apps to api.malibu.tech. Same invite gate.
• End-to-end serving. Live Buyer → gateway → coordinator → Apple Silicon provider → response → signed settlement receipt → settlement. Public snapshot: api.malibu.tech/v1/stats/overview.
• Multi-turn tool calling. Live Standardized tool_calls[] emission. Qwen-family models accept role: "tool" results on later turns; gpt-oss models support the first turn only. Buyer frameworks must validate before execution — Buyer-side validation.
• Structured output. Live response_format: json_schema for a strict schema subset. The provider validates the finished output against your schema; it does not constrain decoding.
• Sticky conversations. Live Send X-MacProvider-Conversation to prefer the provider that served earlier turns. The preference often does not win; when it does, cache hits bill at a lower rate and show in usage.cached_prompt_tokens. Buyers cannot pin a request to a specific provider.
• Signed receipts. Live Most paid requests settle on a signed settlement receipt that the coordinator verifies; the rest carry a legacy signed receipt and settle as quarantined. Some non-streaming responses carry a v0.3 receipt header (currently roughly a quarter to a third); streams never do. For the rest, use /v1/receipts. Pending Offline verification of current receipts is pending a verifier update.
• Verified model settlement. Live Missing, invalid, or late receipts settle with zero provider credit and no buyer debit.
• Rate card & credits. Live Credit-based metering at a 90% provider share today. Payout wallet registration is live; USDC-on-Base payout is planned for public beta. See Economics.
• Model catalog. Live A signed catalog of 17 open-weight MLX models — Llama 3.2 3B and 3.1 8B, Qwen3 / Qwen3.5 / Qwen3.6 / Qwen3.8 and Qwen-Coder rows, Gemma 4, Nemotron 3 Nano, GLM-4.5-Air, gpt-oss-20b and gpt-oss-120b. A model routes only when an eligible Mac with enough memory is online and serving it. Live list: api.malibu.tech/v1/status. Details: Model catalog.
• Routing tiers. Live New Macs join provisional; providers that meet the network’s trust criteria move to trusted and full routing weight. See Payments → Routing tiers.
• Hardware attestation. Beta Apple-backed device attestation is reported for a small part of the pool. It does not affect routing or settlement yet.
• Release cadence. Live Signed provider releases on GitHub with signed autoupdate; public download follows GitHub Latest via malibu.tech/host.

​

Token protocol — Planned

• PoMW mining substrate. Prototype A Metal port of the Proof-of-Model-Walk algorithm passed two hundred verification vectors byte-exact against a Rust reference on Qwen3-1.7B GGUF weights. Internal Phase 1 measurements on an M5 Air (July 2026) — about 2.14 million walks per second at about 4.7 watts, and roughly ten times the per-watt walk rate of an RTX 5090 — are unaudited prototype numbers: the results report is not yet published, and they measure the walk workload, not inference. See Benchmarks & methodology.
• v0 mining launch: Planned v0 on-chain emission, burn-and-mint fee split, token-native settlement.
• v1 marketplace: Planned v1 TOPLOC per-request verification (Prime Intellect lineage) on buyer-paid inference.
• Pooled serving: Planned multi-Mac models beyond one machine’s unified memory.

This is a launch document for a network that already runs inference — and a protocol layer still activating.

​

What Malibu lets you do

Your Mac already has the hardware open models need: unified memory, a fast GPU, and low power draw. Malibu gives that idle capacity a job. The Malibu app installs a provider that serves AI requests to buyers in the background, and you can pause it from the menu bar at any time. A single high-RAM Mac can serve large open models from unified memory. The models that dominate industry inference spend are larger still; pooled serving (several Macs over a local interconnect) is the growth vector for those — on the roadmap, not live. And the bigger the model, the more a verifiable receipt is worth — when you pay frontier rates, you want evidence that the endpoint ran the weights you paid for, not a cheaper stand-in. What a provider earns today is simple to state and impossible to promise: 90% of the credits for the verified inference its Mac serves, at the rates on the public rate card. How much traffic a Mac gets depends on buyer demand, uptime, routing tier, which models it can fit, and thermal limits. Credits accrue on the ledger now; USDC payout is planned for public beta. Malibu publishes no per-Mac earnings figure. The mechanics and the hardware caveats are on Provider economics. Cooperative scheduling keeps the Mac yours. Mining (planned) is designed to run in the background and yield to serving.

​

What buyers get

On comparable OSS SKUs, Malibu is often cheaper than aggregator list — sometimes a lot cheaper on small models, and on some 20B–32B rows even or higher (up to about 59% higher on gpt-oss-20b in the dated snapshot). Dated snapshot: Pricing comparison. Live card: api.malibu.tech/v1/rate-card. Paid requests settle on signed receipts checked by the coordinator. Offline buyer verification of current receipts is pending a verifier update. The signed catalog covers 17 open-weight models, from Llama 3.2 3B up to Qwen3.x, Qwen-Coder, Gemma 4, Nemotron 3 Nano, GLM-4.5-Air and gpt-oss-120b — each runs only when a Mac with enough memory is online and serving it (Model catalog). Drop-in OpenAI-compatible endpoints; in pre-beta, public API use is covered by a free daily token quota. Why buyers switch:

• Cheaper on many OSS rows. Verify against the snapshot; do not assume a flat 20–40% on every SKU.
• Receipt-verified. Signed binding of prompt and output hashes to the provider’s key and its reported model hash — see Signed receipts for what that does and does not prove. Per-request forward-pass attestation via TOPLOC ships at v1 — not live on buyer traffic today.
• OSS-native. Best model per task, when that model is warm. No vendor deprecating a closed SKU next quarter.
• Cooperative trust. Not confidential compute — prompts route through providers as plaintext. See Security & trust model and Threat model.
• Buyer-side validation. Emitted tool_calls[] are model output, not provider-verified intent. See Buyer-side validation.
• No lock-in. Change providers without changing OpenAI-SDK code.

Malibu is designed to be adopted without changing anything on the buyer side except the endpoint.

​

How it works

Every provider runs one binary. At mining launch it performs Proof-of-Model-Walk — a byte-exact 256-step walk through the canonical open-source model’s weights, terminated by a cryptographic commitment. Deterministic. Verifiable in microseconds. The walk hashes over the same weight file the network sells inference for, using the same Metal integer pipeline. Every mined block is a receipt that a specific Mac exercised the memory substrate inference depends on. When a buyer request arrives at v1, the miner yields in ~50 milliseconds — weights already resident, pipeline already running, file already the file the buyer is paying for. Mining state is serving state. That equivalence is the architectural bet the whole network is built on. It is not enforced in production today. Verification is cheap. PoMW verifies in microseconds. TOPLOC (Prime Intellect, 2025) will verify buyer responses at about one percent of serving cost at v1.

​

The token flywheel

Under the planned token economics, every dollar entering Malibu would strengthen the network: providers would get paid, reserves would grow, and tokens would be bought and burned automatically. None of this is live today. When a buyer pays $1 for inference under token economics:

• 70¢ goes to the provider who served the request.
• 12¢ grows a protocol-owned reserves floor — held in USDC, on-chain, backing the token’s minimum value.
• 18¢ automatically buys $MALIBU on the open market and destroys it.

No oracle. No governance vote. No token-holder decision. Every transaction is protocol-enforced $MALIBU demand and protocol-owned reserve growth. The token, minted through work, is burned through use. Provider issuance runs on a deterministic emission curve — front-loaded to bootstrap the fleet, tapering as buyer USDC replaces it.

​

Why Malibu wins

1. Apple already shipped the hardware. Roughly two hundred million Apple Silicon Macs with unified memory. Deployed, paid for, and often idle.
2. PoMW is designed to prepare the fleet before buyers arrive. Providers earn from day one of mining launch for verifiable useful work. That layer is not paying in production yet.
3. Receipts today; TOPLOC at v1. Every paid request settles on a signed receipt now. TOPLOC adds per-request forward-pass attestation when that milestone ships.
4. Every transaction is designed to strengthen the token. At launch, buyer USDC automatically buys and burns $MALIBU; reserves grow. Not active on today’s 90/10 credit rail.
5. The network compounds. Every new provider increases coverage. Better coverage attracts more buyers. Buyer revenue increases provider earnings. Higher earnings attract more providers.
6. The combination is the bet — not a monopoly on the pieces. Akash, Prime Intellect, Render, and io.net already cover overlapping slices (decentralized compute, verification lineage, GPU marketplaces). Malibu’s edge is Apple Silicon + receipts now + PoMW-before-demand as a token mechanism those businesses have little reason to copy. The hardware is not Malibu’s. The mechanism has to come from outside Apple, OpenAI, and the hyperscalers.

​

Join

The hardware already exists. Malibu puts it to work. The live pool is a small pre-beta (7 Macs online on the 23 Sep 2026 snapshot — Network status). Emission economics, when they ship, are designed to pay early miners more; mining reputation is designed to transfer to serving priority at v1. Apple built the hardware. Malibu gives it a job.