Model registry

The model registry is the catalog every node reads to decide what it can serve and every client reads to decide what it can call. It is metadata, not an allowlist: a provider can serve any compatible open-weight model, and the catalog tells the network what each entry is, what it needs and how to verify it. Tenzro ships no first-party models.

Browse the live registry, with the providers currently serving each model and their prices, at /models.

What an entry declares

Each catalog entry carries what a node needs to know before it fetches a single byte:

• Modality: language, embedding, vision, speech, segmentation, detection, forecasting, video or media generation.
• Shape: dense or mixture-of-experts, quantisation, context window, weights size and the memory needed to hold it.
• Content hash: the pinned hash of the weights, checked before any load (see below).
• Licence tier: Permissive, Attribution, CommercialCustom or NonCommercial.
• Serving profile: the recommended sampling settings from the model's authors, applied by default and overridable per request.
• Speculative decoding pair: an optional drafter or built-in Multi-Token Prediction (MTP) head.
• Multimodal projector: for vision-language models, the projector file that lets the model accept images.

Catalog coverage

Language models, dense and mixture-of-experts, from small edge models to very large MoE models:

Qwen          3 / 3.5 / 3.6, including vision-language variants
Gemma         4 (with MTP heads)
DeepSeek      V3 / V4          Kimi       K2 / K3
GLM           5.x              MiniMax    M-series
Nemotron      Nano / Ultra     gpt-oss
Mistral, Phi, Granite, SmolLM and others

Every other modality:

Media generation   text-to-image, image-to-image, text-to-video, image-to-video
Vision             CLIP, SigLIP2, DINOv3
Speech in          Moonshine, Distil-Whisper, Whisper, Parakeet, Canary
Speech out         Qwen3-TTS
Text embedding     Qwen3-Embedding, EmbeddingGemma, BGE-M3, ModernBERT-embed
Segmentation       SAM 2, SAM 3 (open vocabulary), EdgeSAM, MobileSAM
Detection          RF-DETR, D-FINE
Forecasting        Chronos-2, TimesFM 2.5, TiRex

Model names identify open-weight releases by their publishers. Listing a model does not imply any relationship with its publisher.

Hash-verified weights from any origin

Weights are content-addressed. Each catalog entry pins the hash of its weights, and a node verifies what it fetched against that pin before loading it. A mismatch refuses the load.

Because verification does not depend on where the bytes came from, weights can be stored across many origins: peers on the network that already hold the model (fetched over iroh with every block verified in transit), the publisher's own repository, or any mirror. A node tries peers first and publishes what it fetched back to its local store, so the next node can fetch from it.

tenzro model download qwen3-8b
tenzro model get-hash qwen3-8b
tenzro model list-hashes

Over JSON-RPC, tenzro_getModelHash and tenzro_listModelHashes are open reads. A fetched artifact is addressable as tenzro://blob/<hash>. See Model provenance for how weight hashes, serving receipts and certification fit together.

Certification and ratings

Any issuer can certify or rate a model: a lab, an independent evaluator, a regulator, a community. Certifications and ratings are signed credentials that reference the model's content hash, so they stay attached to the exact weights they describe. A relying party chooses which issuers it trusts and filters on their credentials; there is no central registry deciding which models are acceptable. Compliance tiers come from those credentials. See Trust and provenance and Credentials.

Licence tiers

Every entry carries a licence tier, enforced when a node registers the model:

Multi-Token Prediction

Entries can declare a speculative-decoding pair. Generic pairs a smaller vocabulary-matched drafter with the target; DraftMtp uses a jointly trained MTP head that mirrors the target's distribution. Serving a model loads its paired drafter automatically, and a drafter problem never fails the serve. Callers opt in per request with draft_n:

tenzro chat gemma4-12b --draft-n 2

Operators who are short of memory can serve without a drafter using tenzro model serve <id> --no-speculative.

Listing models

tenzro model list
tenzro model info gemma4-12b

The catalog calls describe what the registry carries. The list*Models calls report what a given node has loaded right now.

Next

• Serve a model and earn TNZO: Model serving.
• Call models: Inference and OpenAI-compatible API.