Time for a standard
Publishers can call their models whatever sells them. Nobody is arguing with their marketing choices; one line alongside the name is all it takes to say what the name does not:
| The name | Always the current one | Pinned to this release |
|---|---|---|
| GPT-6.2 Astra | model:openai/text/ultra | model:openai/text/ultra:6.2 |
| Gemini 4 Hydrogen | model:google/text/core | model:google/text/core:4 |
| Claude Haiku 5.5 | model:anthropic/text/light | model:anthropic/text/light:5.5 |
ultra is the flagship, core the everyday default, light the fast, cheap rung. Leave the version out and the line follows the next launch on its own. Pin it only where a result must be reproduced.
Switch provider by changing one word:
model:openai/text/core
model:google/text/core
model:anthropic/text/core
Each line is that publisher's everyday default. Same role, not the same results.
A few things a product name cannot express:
model:anthropic/text/core@bedrock | The same model served by someone else, under the same name. |
model:anthropic/text/ultra-restricted | A model restricted to selected operators, flagged directly in the identifier. |
model:alibaba/text/core-open/35b-a3b/mxfp4@local | The newest open Qwen at the size and precision your GPU fits. |
The moratorium
Nobody has to give up a brand. The moratorium asks for four small, verifiable commitments from publishers and tooling rather than from application developers:
- Publishers: publish a model URI next to the native name in model cards and in the model-listing API response (one extra field).
- Publishers: place every model on the shared six-rung scale at release, with at most one model per slot (rung, scope, suffixes and size). A new model in an occupied slot replaces the old one.
- Publishers: never change the model behind a pinned version. An update gets a new version.
- Routers, gateways and local inference engines: accept a model URI as an alias of the native name.
This is an appeal. By signing, developers ask publishers and tooling to take on these commitments, and adopt model URIs in their own configuration files and documentation, with the native name as a fallback until both sides support them.
The working draft
A model URI says what role a model plays in its publisher's lineup, who may use it, how it is licensed and where it runs. It carries verifiable facts only, with no quality scores or opinions.
model:publisher/scope/tier
[-special][-access][-weights]
[/size][/quant][:version][~derivation][@host]
Written on one line, without spaces. Only the first three parts are required.
| Part | Question it answers | Values |
|---|---|---|
publisher | Who made the model? | A registered organisation name: anthropic, google, alibaba |
scope | What does it produce? | text omni image video audio tts transcribe embed rerank decision |
tier | Which role in the lineup? | nano light core high max ultra |
special | What is it published for? | A registered specialisation: code, rt, base, guard, ocr, research, gui, medical, voiceclone, … |
access | Who may use it? | restricted gov internal |
weights | Can the weights be downloaded? | open unmodified OSI-approved licence, weights any other downloadable licence |
size | How big is it? | Total parameters (m, b, t), plus active ones for MoE: 35b-a3b, 2.4t-a95b |
quant | At what precision does it run? | q4 fp8 mxfp4 |
version | Which release? | Leave it out for the current model, or pin the publisher's number |
derivation | Did a third party modify it, and how? | Author and kind: huihui-abliterated |
host | Where does it run? | local, edge, bedrock, vertex; absent means the publisher's API |
The tier is a role, not a score
| Tier | Role |
|---|---|
nano | The smallest and cheapest, for simple high-volume tasks |
light | Fast and inexpensive |
core | The publisher's everyday default |
high | A step above the default, for harder work |
max | Just below the flagship |
ultra | The publisher's flagship |
The scale is the same for every publisher. A publisher that adopts the standard declares the tier at release and fits its lineup to the scale rather than stretching the scale to fit. Until it does, the registry assigns the tier: core is the model the publisher presents as its everyday default, ultra its flagship, and the others are placed by list price.
Four rules that keep it stable
- One slot, one model. Within a scope, a combination of suffixes and a size, a new model on a rung replaces the old one, so a line without a version always means exactly one model.
ultraandultra-restrictedare two slots. - No version means the current model. With size and precision it follows the newest release at that size and precision. If the next generation drops that size, the line stays where it is and the resolver says so.
- Pinned versions never change. No silent updates, so old logs, configs and benchmarks stay reproducible.
- The model is separate from where it runs. Who built the weights and who hosts them are independent.
Examples
| Model | Model URI |
|---|---|
| GPT-6.2 Astra | model:openai/text/ultrapinned: model:openai/text/ultra:6.2 |
| GPT-6 Luna | model:openai/text/nanopinned: model:openai/text/nano:6 |
| Gemini 4 Carbon | model:google/text/ultrapinned: model:google/text/ultra:4 |
| Claude Sonnet 5.5 on Bedrock | model:anthropic/text/core@bedrockpinned: model:anthropic/text/core:5.5@bedrock |
| Claude Mythos 5.2, selected operators only | model:anthropic/text/ultra-restricted:5.2 |
| A government-only edition | model:example/text/core-gov:2.0 |
| Mistral Codestral | model:mistral/text/core-codepinned: model:mistral/text/core-code:25.08 |
| Gemini 2.5 Computer Use (GUI & browser control) | model:google/text/core-guipinned: model:google/text/core-gui:2.5 |
| Gemini Deep Research Preview | model:google/text/core-researchpinned: model:google/text/core-research:2026.04 |
| Baichuan M2, medical, 32B | model:baichuan/text/core-medical-open/32b@localpinned: model:baichuan/text/core-medical-open/32b:2@local |
| Qwen 3.6 MoE, 35B total, 3B active, MXFP4 | model:alibaba/text/core-open/35b-a3b/mxfp4@localpinned: model:alibaba/text/core-open/35b-a3b/mxfp4:3.6@local |
| The same model, abliterated by a third party | model:alibaba/text/core-open/35b-a3b/q4:3.6~huihui-abliterated@local |
| Gemma 4, 26B total, 4B active, Q4 (Apache-2.0) | model:google/text/core-open/26b-a4b/q4@localpinned: model:google/text/core-open/26b-a4b/q4:4@local |
| Meta Llama 4 Maverick (community licence; one size on this rung) | model:meta/text/core-weights@localpinned: model:meta/text/core-weights:4@local |
| GPT Realtime 2.1 (real-time duplex voice) | model:openai/omni/core-rtpinned: model:openai/omni/core-rt:2.1 |
| MiMo-V2.5-TTS VoiceClone | model:xiaomi/tts/core-voiceclonepinned: model:xiaomi/tts/core-voiceclone:2.5 |
| Lyria 3 Pro Preview (music & non-speech audio) | model:google/audio/corepinned: model:google/audio/core:3 |
| Liquid d1 (typed probabilistic decision) | model:liquid/decision/corepinned: model:liquid/decision/core:1 |
Tiers are assigned with the registry rule from the publishers' own descriptions and price lists, as of 21 November 2026. No publisher has declared them yet. More examples are in SPEC.md.
What it is not
- Not a benchmark. Two models on the same rung fill the same role, though their quality and behaviour still differ.
- Not a download URL. A model URI identifies a model; fetching weights is left to existing registries and hosts.
- Not a catalogue. Context windows, prices and input modalities belong in the registry or in existing catalogues.
The full draft has the grammar, a reference regex, the registry format, security considerations and the open questions: read the specification.
The repository already includes a seed registry built from models.dev (over 400 models from more than 60 publishers), a list of incompatible catalogue entries (missing evidence, rolling aliases, slot collisions, parameter splits and opaque sizes), a small Node CLI resolver and generated LiteLLM aliases.
Sign, or break it
This is a working draft, not a finished standard. It gets better when people attack it.
To sign, add one line with your name, role and an optional link to SIGNATORIES.md. To object, open an issue. An edge case the grammar cannot express is the most useful report you can send.