​
Model catalog

The set of models the network knows how to serve, and how the catalog is maintained.
Three lists, three jobs:

​
Catalog (eligible — runs when matching RAM is online)

The network is built to serve these families. A model routes when at least one admitted Mac has it warm. Higher-RAM Macs (Studio / 32 GB+) bring 35B–120B class online; they are not “future models.” Warm vs priced vs catalog are still three lists. Call GET /v1/models (Bearer) before you depend on a specific ID in production.

​
Warm now (16 August 2026)

The live pool was small. Auth’d GET /v1/models returned: That set grows as higher-RAM Macs join and load catalog weights.

​
Priced rows (rate card)

Live card: GET https://api.malibu.tech/v1/rate-card (generated_at 2026-07-29, usd_per_million_credits = 1). Same payload on https://malibu.tech/v1/rate-card. Rate-card IDs are slugs. Gateway chat uses MLX serve IDs. Map them yourself; they are not always identical strings. Default completion rate is $1.00 / 1M (1,000,000 credits). Provider share on every row is 90% (provider_share_bps: 9000). Catalog families without a named row (Qwen2.5 7B, Llama 3.3 70B, gpt-oss 120B, Qwen3.5 35B) still bill at default until a dedicated row is published — they remain catalog-eligible.

​
What the catalog contains

For each catalog entry:

​
Catalog distribution

The catalog is served as a signed static feed off the release infrastructure. Providers pull it on install and on autoupdate. Signature verification anchors:
  • Which models are eligible to serve.
  • Which weight hashes count as canonical (receipts referencing an off-catalog model_hash verify invalid).
  • Which templates are marked tool-calling / structured-output capable.

​
Adding a model

Adding a new model to the catalog is an operator action. Steps:
  1. Publish MLX weight files with a canonical SHA-256.
  2. Add a catalog entry with model_id, model_hash, RAM tier, and rate card.
  3. Sign and publish the updated catalog feed.
  4. Existing providers pick up the update on next autoupdate cycle; new installs get it directly.
Community-proposed models: file an issue on MalibuAI/malibu with the model ID, weight file, and target RAM tier.

​
Model retirement

Retiring a model means:
  • Removing it from the signed catalog feed.
  • Providers stop advertising it warm on next feed refresh.
  • Buyer requests for a retired model return 404 model_not_available.
Receipts issued while the model was in the catalog remain verifiable indefinitely.

​
Receipt binding

Receipts carry model_hash, and malibu-verify 1.1.x cross-checks it against the catalog. A valid receipt therefore certifies:
  • The signed tuple was produced by the provider’s key.
  • The provider claimed to run this specific model file (by hash).
  • That file is in the catalog.
Full trust boundary: Security & trust model.