models · open weights
The Qwen3.8 family, the baseline.
An open-weights family in NVFP4 that fits two workstation cards. It is the cheapest cell in the catalog — the one that proves the whole path for pocket change before you point it at a bigger checkpoint.
| weights | NVFP4 — the canonical two-card pack; 126 GiB |
|---|---|
| context | 262,144 tokens configured |
| hardware | 2× RTX PRO 6000 (Blackwell SM120, 96 GB each) |
| tensor parallel | 2 |
| engine | SGLang with EAGLE 3/1/4 decode and the linear-replay SSM speculative path |
| observed price | $2.94 — an hour, observed; the class band is $2.14–$3.47 |
| catalog id | RadixArk/Qwen3.8-Flash-Next-NVFP4 |
The one TP2 shape that actually fits.
126 GiB over two cards is 63 GiB per card, under the 67 GiB ceiling. That is the entire reason this pack exists: the official FP8 build cannot tensor-parallel 2 at all, because its mixture-of-experts intermediate size halves from 640 to 320 and 320 is not divisible by the 128×128 FP8 weight block. The weight creation raises before any card is warm.
Expert parallelism shards each expert whole and is the four-card cell's fix for the FP8 build. The NVFP4 pack needs no such fix, and it is the only shape in the catalog where two cards genuinely serve.
The PLE embedding table never fits a 96 GiB card: it is offloaded to host RAM at roughly 51 GiB, which is what the 200 GiB host floor is for. And the QSA sparse-attention gate needs a prelude patch on the image — without it, the backend refuses the SM120 part.
The cheap cell is a proof, not a toy.
Under four-way concurrency each stream held about 63 tokens a second at roughly 430 ms to first token on the live two-card cell. That is a real serving shape, not a smoke test — and at $2.94/h observed it costs less than a coffee to run the whole path end to end. Run it before you trust the kit with a 290 GiB package.
The numbers this shape is known by.
63 GiB weights per card at TP2 194 t/s decode, four concurrent streams 90 ms time to first token, measured $2.94 an hour, observed Run it.
The profile is in the kit already. This is the exact command, and the fit gate runs inside
it before anything is created.
the snipe, by model name
python3 package/launch/snipe.py --model RadixArk/Qwen3.8-Flash-Next-NVFP4
or ask the catalog first, with no key and no billing
./target/debug/onefrontier fit RadixArk/Qwen3.8-Flash-Next-NVFP4 --tp 2
./target/debug/onefrontier best RadixArk/Qwen3.8-Flash-Next-NVFP4
fit exits non-zero on a halt, so a script gates on the status
directly. best ranks the classes the gate passes by cheapest
observed band and names the ones it rejects, with the reason.
one command, whatever the model
Serve it on hardware you can rent this afternoon.
The gate stops you before the paid clock, the tunnel arms on boot, and the flags come from
someone else's paid hour.