model-catalog

Model catalog data as JSON files, refreshed on a schedule.

https://raw.githubusercontent.com/cloudstack-llc/model-catalog/main/v1/prices.json
https://raw.githubusercontent.com/cloudstack-llc/model-catalog/main/v1/ollama-models.json
https://raw.githubusercontent.com/cloudstack-llc/model-catalog/main/v1/featured-models.json
https://raw.githubusercontent.com/cloudstack-llc/model-catalog/main/v1/smart-route-starters.json

File	Contents	Source	Cadence
v1/prices.json	Token pricing for hosted models	models.dev	6 hours
v1/ollama-models.json	The Ollama library: models, tags, sizes, context windows	ollama.com	12 hours
v1/featured-models.json	Curated local models worth downloading, grouped into collections	Curated; resolved against Ollama and Hugging Face	Daily
v1/smart-route-starters.json	Model-neutral Smart Route starters, grouped by use	Curated	On change
v2/runtime/{stable,dev}/runtime-release-feed.signed.json	Signed Msty Nexus Runtime releases	Ollama, llama.cpp, and Msty MLX metadata	4 hours

Token pricing

Format

Every model with a published input and output rate, in US dollars per one million tokens, grouped by provider.

"anthropic": {
  "claude-opus-4-5": {"input":5,"output":25,"cache_read":0.5,"cache_write":6.25,"context":200000,"max_output":64000}
}

Field	Meaning
input, output	Required. USD per million tokens.
cache_read, cache_write	Prompt-cache rates, when published.
reasoning, input_audio, output_audio	Modality-specific rates, when published.
context, max_output	Token limits.
tiers	Long-context pricing, ascending by above_context. Rates replace the base rates once a request's context exceeds that size.

An absent field means the rate is not published — never zero. Models with no price at all are omitted rather than published as free.

The envelope carries schema_version, generated_at, counts, and the upstream etag and sha256.

v1/ is versioned on purpose. A breaking schema change ships as v2/ alongside it.

Refresh

A scheduled job regenerates the file every six hours and commits it only when the prices changed, so git log -p v1/prices.json shows actual price movement.

A run that finds no priced models, or fewer than 70% of the previous run's, fails without writing. Individual models that fail validation are dropped rather than repaired. CI re-serializes the committed file and range-checks every rate.

These are estimates. Providers change prices without notice and negotiated rates differ from list rates. Bill against your provider's invoice.

Development

node --test scripts/*.test.mjs   # transform rules
node scripts/verify.mjs          # validate the committed file, no network
node scripts/generate.mjs        # fetch upstream and rewrite v1/prices.json

No dependencies.

Runtime release feeds

The runtime catalog lets Msty Nexus update managed runtimes independently from Nexus app releases. Every four hours, the publisher checks stable Ollama and llama.cpp releases plus the latest reviewed Msty MLX runtime. It writes only when a runtime version or immutable artifact identity changes.

The signed schema-2 feeds are served directly from this public repository:

https://raw.githubusercontent.com/cloudstack-llc/catalogs/main/v2/runtime/stable/runtime-release-feed.signed.json
https://raw.githubusercontent.com/cloudstack-llc/catalogs/main/v2/runtime/dev/runtime-release-feed.signed.json

They are not mirrored to R2. The signature is the trust boundary; the hosting origin supplies bytes but cannot authorize a runtime artifact.

Direct GitHub release assets are not downloaded by the publisher. GitHub's SHA-256 digest and byte size enter the signed feed, and Nexus verifies the download against that signed digest during installation. The Msty MLX archive uses its published checksums.txt and an HTTP metadata request for size.

Upstream llama.cpp currently splits the Windows CUDA package into the server and CUDA runtime ZIPs. For that variant only, the publisher downloads both digest-pinned components, rejects unsafe or duplicate paths, merges them once, and publishes the result as an immutable GitHub Release asset. Later refreshes reuse that asset.

The feed uses the existing backend-aware Nexus schema and Ed25519 trust anchor. The signing seed is stored only in the MSTY_NEXUS_RUNTIME_RELEASE_FEED_PRIVATE_KEY repository secret. Pull requests never receive it. REFRESH_TOKEN remains the repository-scoped token used by the other scheduled catalogs; the default Actions token is a functional fallback.

node --test scripts/runtime-feed.test.mjs
python3 -m unittest discover -s scripts -p '*_test.py'
node scripts/runtime-feed-verify-all.mjs

Smart Route starters

v1/smart-route-starters.json gives Msty Nexus a model-neutral starting point for common routing jobs. A starter contains the route name, classifier lanes, target guidance, and the endpoint families it expects. It never contains a model or pool ID. Nexus asks the person adding it to choose those targets.

Collections reference starters by ID, so one starter can appear in more than one part of the catalog without being copied. revision changes when the curated starter changes. Existing Smart Routes are independent copies and are never changed by a catalog update.

Displayed copy describes each starter and lane directly. Lane descriptions are capped at 512 bytes because that is the classifier input limit.

Writing lane descriptions

Arch-Router matches the latest request against every lane as a natural-language route policy. Its training uses a domain-action taxonomy: the starter supplies the domain, while each lane should name a distinct action within that domain. See the model card and paper.

• Separate lanes by requested action, not by vague model qualities such as "easy" and "strong."
• Include concrete request signals, source material, and expected artifacts.
• State a boundary when two neighboring lanes could plausibly match the same opening request.
• Write each description for its actual job. Do not force every lane through one sentence template.
• Review the five lanes as a set and test short, ambiguous, and follow-up requests before publishing.

The fallback handles requests that do not clearly match a lane.

Validate a catalog edit before committing it:

node scripts/smart-route-starters-verify.mjs

Optional REFRESH_TOKEN secret: a fine-grained PAT with contents: write on this repository. GitHub disables scheduled workflows after 60 days without repository activity, and pushes made with the default GITHUB_TOKEN do not reset that clock.

License

MIT (see LICENSE). Data derives from models.dev, also MIT — see NOTICE.

Ollama library

v1/ollama-models.json — every model in the Ollama library, with its tags, download sizes, context windows, and per-tag parameter counts and quantization.

{
  "name": "gpt-oss",
  "description": "OpenAI's open-weight models ...",
  "tags": [
    {"tag":"20b","size":"14GB","digest":"17052f91a42e",
     "model_info":{"contextWindow":"128K","parameters":"20.9b","quantization":"MXFP4","arch":"gptoss"}}
  ],
  "params": ["tools","thinking","cloud","20b","120b"],
  "pulls": "11.8M",
  "pulls_approx": 11800000,
  "updated": "1 month ago",
  "updated_at": "2026-07-15T18:02:00Z"
}

name, description, tags, params, pulls, and updated are present on every model. digest, arch, pulls_approx, updated_at, cloud, cloud_tags, and model_info.projector are additions.

pulls and updated are true only at generated_at. They are the strings the site renders, so a file written at midnight still says "1 month ago" the next evening. updated_at carries the absolute timestamp the relative string is derived from, and pulls_approx decodes the display value so it can be sorted — it is a decoded approximation, not a true count, because no absolute pull count is published anywhere. Change detection ignores all four.

An unknown context window is "N/A", never an empty string.

Some tags carry an empty parameters and quantization. Ollama renders no metadata block at all for certain variants (-mlx, -mxfp8, -nvfp4), so the values do not exist to be scraped; those tags still carry a real size and contextWindow. The count is declared in counts.missing_model_info and checked in CI, so a parser regression cannot hide among them.

Cloud tags are excluded from tags[]. They host no weights, so they have no size, parameters, or quantization. A model serving them carries "cloud": true and lists them in cloud_tags.

Vision models keep their projector separate. The library renders two metadata blocks for them — the model and its CLIP projector. model_info holds the model; the projector is under model_info.projector.

Source

Ollama publishes no API. Every API-shaped path returns 404, content negotiation is ignored, and ollama.com/api/tags returns hosted cloud models rather than the library. The catalog is therefore built from server-rendered HTML: one request for the library index, one per model for its tags, and one per distinct layer digest for the fields that appear nowhere else.

Parameter counts and quantization are immutable for a layer digest, so they are cached in v1/ollama-model-info-cache.json. A cold run is ~6,500 requests; a warm one is a few hundred. The crawl runs 8 concurrent with a truthful, contactable user agent.

Because this rests on presentation markup, the generator refuses to publish when a field drops below 95% coverage, when the catalog shrinks by more than 10%, or when tag names and scraped parameter counts start disagreeing. On refusal the job fails and the previous file stands.

Featured models

v1/featured-models.json — the local models worth recommending to somebody who has just installed a runtime and does not yet know what to download.

Unlike the scraped files this one is curated. Which models appear is a judgement, and it lives in scripts/featured-curation.mjs; every value attached to them is not. Tags, sizes, context windows and parameter counts come from ollama-models.json, and repositories, GGUF file names and byte counts come from the Hugging Face API. A reference that cannot be confirmed fails the build rather than shipping — a featured row with a dead download behind it is worse than one fewer row.

Format

A featured entry is a model, not a repository. One entry carries every way to obtain that model, so a client can offer "Qwen3.5 9B" and let the engine and quantization be a detail behind it. Collections reference models by id rather than nesting them, so one model can appear in two collections without being described twice.

{
  "schema_version": 1,
  "generated_at": "2026-08-15T00:00:00Z",
  "collections": [
    {"id":"start-here","title":"Start here","description":"...","models":["qwen3.5-9b"]}
  ],
  "models": [
    {
      "id": "qwen3.5-9b",
      "name": "Qwen3.5 9B",
      "publisher": "Qwen",
      "summary": "Strong all-round chat with a very long context.",
      "parameters": "9.7B",
      "context": 262144,
      "capabilities": ["chat","tools","reasoning","vision"],
      "installs": [
        {"runtime":"ollama","model":"qwen3.5:9b","quantization":"Q4_K_M","size_bytes":6600000000},
        {"runtime":"llamacpp","repository":"unsloth/Qwen3.5-9B-GGUF","file":"Qwen3.5-9B-Q4_K_M.gguf",
         "quantization":"Q4_K_M","size_bytes":7500000000,"projector":"mmproj-F32.gguf"},
        {"runtime":"mlx","repository":"mlx-community/Qwen3.5-9B-MLX-4bit","size_bytes":6000000000}
      ]
    }
  ]
}

Field	Meaning
id, name	Required. Everything else is optional.
publisher, summary, parameters, context, capabilities, tags	Description, for display and filtering. "coding" in tags marks a model that is good at writing code.
tier	Quality tier: frontier, excellent, strong, or good.
installs[].runtime	ollama, llamacpp, or mlx.
installs[].model	The Ollama tag. Ollama installs are addressed by it; Hugging Face installs are not.
installs[].repository, file, revision	The Hugging Face locator. revision defaults to main.
installs[].quantization, size_bytes	What the download is and how big.
installs[].projector	A file name in the same repository.

size_bytes is what the download costs, so a llama.cpp entry with a projector counts both files. The order of installs is the recommendation order: a consumer offers the first entry it can actually run.

projector is llama.cpp only. A GGUF vision model needs its mmproj-* file loaded beside the weights, and a model served without one answers questions about images it cannot see. Ollama and MLX ship the projector inside the artifact they already download, so they never name it. A publisher shipping a projector is also treated as stating the model is multimodal, so those entries declare vision.

Only single-file GGUFs are featured. A featured row promises one click, and a sharded download is a flow this format cannot describe. A repository publishing only shards gets no llamacpp entry.

tier is a curated quality rating. Msty Nexus Cluster Advisor ranks by it. It rates general chat and reasoning quality against the open models available at the time:

Tier	Meaning
frontier	The best open models, close to hosted frontier models.
excellent	One step below: the previous generation's flagships and the top coding models.
strong	Capable mid-size models, roughly 20B to 130B.
good	Small models, and models with one narrow job (OCR, embeddings).

Tiers are relative, so review them when a new generation of models is released. A client that finds no tier estimates one from the model's size and says so. The field is additive: older Nexus builds ignore it, so schema_version stays 1.

The cluster collection is a contract. Cluster Advisor reads it by id. It holds MLX models too large for one Mac, from model families the cluster runtime can split across Macs.

An Ollama size published as a range uses the upper end. Ollama shows a range, such as 6.6GB - 7.6GB, for a tag that resolves to different files per platform. size_bytes is what the download may cost, so a fit check that overestimates is the safer error.

Refresh

A daily job (refresh-featured.yml) rebuilds the file from the curation, the committed ollama-models.json, and the Hugging Face API. It refreshes files, sizes, parameters, context and popularity, and never adds or drops a model: only a reviewed edit of scripts/featured-curation.mjs does that.

The job commits only when something other than generated_at or popularity changed; Msty Nexus reads popularity live and treats the published numbers as a fallback. Any reference it cannot confirm, or any Ollama size it cannot read, fails the run without writing, so the published file stands until a maintainer fixes the curation.

After each run, featured-report.mjs lists mlx-community repositories that may be newer versions of featured models (by each model's watch term) in the job summary. These are leads for a maintainer, not changes.

To change the list, edit scripts/featured-curation.mjs, then rebuild and commit both:

node scripts/featured-generate.mjs   # fetch and rewrite v1/featured-models.json
node scripts/featured-verify.mjs     # offline: shape, sizes, and match with the curation
node --test scripts/featured.test.mjs

CI runs the verifier, so a curation edit without a rebuild fails the pull request.