Changelog
Every major change to the catalog and resolver, newest first. Browse the catalog to see the results.
The Workbench used to work from a small demo list. It now loads the real published catalog when the page opens — 110 models, 40 harnesses and 20 gateways — so lookups and refusals reflect what is actually on record.
Feedback form submissions already became issues in the public HiQS-AI-Resolve-Catalog repository. After submitting, the confirmation now shows a direct link to your issue, so you can follow it without an account.
model:stealth/space-bunny-alpha@r1 is on record as an anonymous stealth preview: fast inference, strong coding, native multimodal input, adjustable reasoning effort, 1M-token context, free during preview. All details are operator-recorded from the OpenRouter page, not provider-verified. Like the other stealth entries, it points to the revealed model once an identity is disclosed.
The site now has a changelog page — this one — with every major change since the beta launch.
Probes, Methodology and Guarantees moved into a Details dropdown in the navigation to keep the top bar short.
The homepage now says “Use at your own risk” everywhere it previously hedged about production trust, and the command box explains that local runtimes (Ollama, LM Studio, MLX) are discoverable records — the resolver routes hosted traffic only.
The ai-catalog-update skill now ships with everything it needs to run a daily catalog session on a local Mac: grounded discovery, duplicate checks against namespaces and source links, and a bundled resolve job script.
It drafts only. Every entry it finds still waits for a human operator to review and publish.
The skill is published in the public HiQS-AI-Resolve-Catalog repository under skills/ai-catalog-update.
runtime:ddalcu/mlx-serve@r1 is a native Apple Silicon server that runs MLX and GGUF models behind OpenAI-, Anthropic- and Ollama-compatible APIs. MIT licensed.
Source badges now say “unverified” when no one has checked them, and entry notes are shown on every page. The homepage now leads with a plain question, its example commands point at real open-weight entries, and a real NO_SERVING_BINDING refusal replaces the synthetic demo.
MLX LM (ml-explore/mlx-lm) was added as the first Apple-native runtime.
The four newest published entries now appear below the hero, each with its one-line description and key details.
The anonymous read endpoint is now a one-click copy command in the hero and in a new footer on every page. Beta wording changed to “Use at your own risk.”
Entries can now record the model they were derived from, as the publisher writes it, plus a neutral one-sentence description. Every existing entry was backfilled from facts already on record.
Entries gained a general source link and license fields alongside the Hugging Face link. A daily job now saves a private ZIP backup of the whole catalog, kept for 30 days.
The part before the slash is the lab’s own short name, and the part after is the vendor’s exact published identifier. The whole catalog was audited against it, and the Google entries were consolidated.
When a stealth preview is later revealed (Ox Alpha → GLM-5.3-Flash, Union Alpha → Pareto), the preview entry keeps its history and points to the revealed model instead of being renamed.
Anyone can now read published entries without an account at resolve.hiqs.ai/api/public/registry. Snapshots are mirrored to the public HiQS-AI-Resolve-Catalog repository.
Local inference engines (Ollama, LM Studio, vLLM, llama.cpp, SGLang) now have their own kind. They can be found in the catalog but don’t take part in resolving a route.
Maintainers can run capped probe matrices whose receipts cannot be edited, and review candidate attestations in one step that is recorded in the audit trail.
The registry, public catalog, entry pages, workbench, and CLI went live as an unsigned beta, with Google sign-in and a single writer path for the site, API, and command line.