the launch card · one per HF weight path
The gap is between the weights and the GPU.
A Hugging Face repo tells you what a model weighs. It does not tell you which quant ships, which engine boots it, what the argv is, or what box it needs. That decision is a file — and it belongs at the model's own path.
The path is the identity.
A card binds to the weight path and shadows it. Nothing here is a new namespace — it is the namespace Hugging Face already owns, with the missing half of the fact added.
the weights, where they live
huggingface.co/{org}/{model}The checkpoint, the license, the weight bytes. Authoritative, and the only thing it answers is what is this.
the card, at the same path
onefrontier.ai/{org}/{model}
onefrontier.ai/{org}/{model}/performanceThe half HF does not carry: every checkpoint, the exact argv per engine, and every hardware floor. The file is viewable and downloadable; the performance page is the measured cells for that path.
The widget, made into a file.
The vLLM and SGLang recipe sites both ship a Deploy widget: pick a hardware selection, get a command. That command exists only in a browser, only for the selections the site offers, and leaves no record of whether it worked.
The card is the same information as a file you can download, diff, and cite — with the hardware floors as data rather than as a paragraph, and with the launches that do not exist named as gaps instead of omitted.
Its intended consumer is one: point it at a card and it follows the
human-readable launch. When the launch fails, the failure is classified, exactly one
field changes, and it tries again — and either way the outcome is recorded. What works
and what does not is the product's memory.
Four sections, one Dockerfile intuition.
A Dockerfile resolves (base image → build steps → runtime). A card resolves (weights → checkpoint → env → hardware). Same discipline, different axis: the card moves down the stack toward the metal.
20 cards, ordered the way the market is.
By Hugging Face downloads at capture — the popularity order the sweep was run in. Every row links to its card at the HF path.
| weight path | params | weights | ckpts | launches | engines | validated on | gaps |
|---|---|---|---|---|---|---|---|
openai-community/gpt2 | 0.14B | 1.0 GiB | 1 | none | — | — | 3 |
Qwen/Qwen3-0.6B | 0.75B | 1.4 GiB | 1 | none | — | — | 3 |
Qwen/Qwen3-14B | 14.77B | 27.5 GiB | 1 | 1 | vllm | — | 2 |
Qwen/Qwen3-4B | 4.02B | 7.5 GiB | 1 | 1 | vllm | — | 2 |
google/gemma-4-31B-it | 31.27B | 58.3 GiB | 1 | 1 | vllm | h100, mi300x, mi325x, mi355x | 2 |
Qwen/Qwen3-8B | 8.19B | 15.3 GiB | 1 | 1 | vllm | — | 2 |
ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF | — | 274.6 GiB | 5 | none | — | — | 3 |
Qwen/Qwen3.8-27B | 27.78B | 51.7 GiB | 1 | 1 | vllm | b200, gb300, rtx_5090, rtx_5090_2x, rtx_pro_6000 | 2 |
Qwen/Qwen3.5-4B | 4.66B | 8.7 GiB | 1 | 1 | vllm | — | 2 |
Qwen/Qwen3-VL-8B-Instruct | 8.77B | 16.3 GiB | 1 | none | — | — | 3 |
Qwen/Qwen3.5-9B | 9.65B | 18.0 GiB | 1 | 1 | vllm | — | 2 |
deepseek-ai/DeepSeek-V3.2 | 685.36B | 642.1 GiB | 1 | 1 | vllm | h100, h200 | 1 |
zai-org/GLM-5.3-Flash | 321.32B | 305.8 GiB | 1 | 7 | vllm, sglang | b200, gb200, h100, mi355x | 1 |
google/gemma-4-26B-A4B-it | 25.81B | 48.1 GiB | 1 | 1 | vllm | h100, h200, mi300x, mi325x, mi355x, rtx_5090, rtx_pro_6000 | 2 |
openai/gpt-oss-20b | 20.91B | 38.4 GiB | 3 | 1 | vllm | b200, h100, h200, mi300x, mi325x, mi355x | 2 |
Qwen/Qwen2.5-VL-7B-Instruct | 8.29B | 15.4 GiB | 1 | 1 | vllm | — | 2 |
Qwen/Qwen2.5-7B-Instruct | 7.62B | 14.2 GiB | 1 | none | — | — | 3 |
ornith-ai/Ornith-1.5-9B-GGUF | — | 45.7 GiB | 1 | none | — | — | 3 |
Qwen/Qwen2.5-0.5B-Instruct | 0.49B | 0.9 GiB | 1 | none | — | — | 3 |
dphn/dolphin-2.9.1-yi-1.5-34b | 34.39B | 64.1 GiB | 1 | none | — | — | 3 |
A gap is the work queue.
12 of 20 paths have a published launch. The other 8 have none — from the two most-downloaded weights on Hugging Face down. These are launches nobody has done, not models nobody can run.
openai-community/gpt2 27.8M downloads Qwen/Qwen3-0.6B 23.8M downloads ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF 7.8M downloads Qwen/Qwen3-VL-8B-Instruct 6.9M downloads Qwen/Qwen2.5-7B-Instruct 5.4M downloads ornith-ai/Ornith-1.5-9B-GGUF 5.4M downloads Qwen/Qwen2.5-0.5B-Instruct 5.1M downloads dphn/dolphin-2.9.1-yi-1.5-34b 4.8M downloads What the card is not.
It is not a benchmark, and it is not a promise. Every number on a card was copied from a source that published it, and every field the source did not answer is written down as a gap.
Not a guess dressed as a recommendation.
Where vLLM verified a model only on xeon6 — a CPU — the card says so
instead of inventing a GPU row. 6 cards in this sweep carry that exact shape.
Not a replacement for the recipe site.
The card cites it. envs[].source is the URL the command was read from, at
the revision it was read at, so a reader can check it and a drift is visible.
Not per provider.
A card describes the weight path, not the market. Which box you rent, and what it costs, is a measurement — that lives on the performance page, one row per cell.
Not finished, and it says so.
46 named gaps across 20 cards. Each one is a field a source did not answer, kept as a field rather than silently dropped.