​
Benchmarks & methodology

Evidence behind hardware and mining claims in the Litepaper. Last reviewed: 23 September 2026.
Malibu publishes two different performance stories. They must not be conflated: A strong PoMW walk result does not automatically imply the same per-watt advantage on inference TPS. This page scopes the published PoMW numbers.

​
PoMW Phase 1 benchmark (M5 Air)

Source: internal Phase 1 Metal port run (results captured July 2026 on an M5 Air). The plan it followed is mining-prototype-phase1.md. The results report itself and the raw benchmark JSON are not yet published, so treat these numbers as unaudited internal measurements until they are.

​
What this proves

  • The PoMW algorithm runs correctly and efficiently on Apple Silicon Metal.
  • The walk is memory-substrate bound — the workload Malibu targets for mining readiness.
  • Byte-exact determinism against a reference implementation is reproducible.

​
What this does not prove

  • Inference tokens/sec on buyer workloads.
  • That every Mac tier sustains the same thermal profile (fanless Air throttles under long runs).
  • Cross-hardware superiority on workloads other than this specific walk.

​
10× per-watt vs RTX 5090 (scoped claim)

The Litepaper cites roughly 10× better per-watt efficiency vs an RTX 5090 on PoMW memory-hard weight walks only — not on inference TPS. Methodology note: GPU numbers use the same walk algorithm ported to the comparison substrate where available. The RTX 5090 measurement and its methodology are not yet published, so this ratio is directional, not audited. This is a mining-substrate comparison, not a buyer inference benchmark. Do not extrapolate to chat-completion pricing or latency.

​
Inference benchmarks (not yet published)

Malibu does not publish a cross-tier inference TPS table on /docs. What exists today:
  • Console bill line — client-measured decode tok/s for a single stream (not network-wide).
  • Public stats — GET https://api.malibu.tech/v1/stats/overview has volume, pool size, estimated capacity and utilization, plus network-wide tokens per minute over the last 30 minutes (timeseries.tpm_30m). There is no per-provider tokens/sec figure.
  • Pricing — GET https://api.malibu.tech/v1/rate-card
Routing picks any eligible provider, so single-request latency varies with which Mac serves you. When a hardware-tier TPS table ships, it will appear here with model, context length, and date.

​
Market sizing claims

Treat fleet-utilization numbers as planning inputs, not audited network telemetry.

​
Reproduce PoMW results

The Phase 1 plan lives in mining-prototype-phase1.md. The harness and results are not public yet. When the miner probe repository ships, rerun the benchmark harness on your Mac class and compare against the M5 Air baseline above.