Astromesh Nebula

Where models are born

Nebula is the open-model foundry of the Astromesh ecosystem. Every model in the Models catalog is trained, gated, and published here before the runtime ever serves it.

Astromesh Nebula is the foundry that produces the ecosystem’s own models. While the rest of the platform runs and deploys agents, Nebula is the upstream step that builds the open models those agents can route to — fine-tuned, quantized, evaluated behind a quality gate, and shipped to the Hugging Face Hub with a machine-readable contract.

Its first family is Centinela preview — Spanish-first models for finance and back-office work in LATAM. The first release, Centinela-Qwen3-4B, is a financial sentiment classifier (positivo / neutral / negativo) QLoRA-fine-tuned from Qwen/Qwen3-4B, Apache-2.0, cheap enough to self-host even on CPU via GGUF.

What Nebula Is

• An open-model foundry: a reproducible pipeline that turns a base model + a dataset into a published, versioned, evaluated model.
• A quality gate: no model is published unless it clears a tier gate (per-metric thresholds, and must strictly beat the base model).
• A machine contract producer: it compiles a catalog.lock.json that the runtime and the docs consume — labels, output validation, aliases, revisions, and eval metrics.
• GitOps-native: model promotion flows through pull requests, from the foundry all the way to a live serving endpoint.

What Nebula Is NOT

• It is not the agent runtime. Agents execute in Astromesh Core / Node, not in Nebula.
• It does not serve inference. Nebula publishes model artifacts to the Hub; a live endpoint (or local Ollama/GGUF) serves them. See Provider Configuration.
• It is not a general training platform. It is an opinionated pipeline for the ecosystem’s own models, one vertical family at a time.

Ecosystem Role

Nebula sits upstream of everything else: it is the source of the models the runtime routes to. The rest of the ecosystem consumes its output — the compiled catalog and the published Hub repos.

flowchart LR
  subgraph nebula["NEBULA (foundry)"]
    base["base model + dataset"]
    hub["Hugging Face Hub repo"]
    lock["catalog.lock.json"]
    base -- "train/merge · quantize/eval" --> hub
  end
  subgraph astromesh["ASTROMESH (runtime + ops)"]
    runtime["Core Runtime"]
    provider["CentinelaProvider (live endpoint / GGUF)"]
    providers["providers.centinela.yaml"]
    models["Models docs + bindings"]
    endpoint["endpoint"]
    runtime --> provider
    providers -- providerRef --> runtime
    models --> providers
    models --> endpoint
  end
  hub --> providers
  lock --> models

Where to Next

I want to…	Go to
See how a model is actually built	The Foundry Pipeline
Understand the catalog contract & GitOps flow	Catalog & GitOps
Follow a concrete model end-to-end	How a model is made: Centinela
Browse the published model cards	Models