models · the frontier
GLM-5.3, on datacenter silicon.
The full-weight frontier checkpoint: NVFP4, 433 GiB, sized for datacenter cards rather than a workstation. Four or eight of them, and a host memory tier that is per rank, not per box.
| weights | NVFP4; 433 GiB |
|---|---|
| context | 700,000 tokens configured; the row pins the window per class |
| hardware | B200 ×4 or ×8 · B300 ×4 · RTX PRO 6000 ×8 · H200 ×8 |
| tensor parallel | 4 or 8, pinned per class |
| engine | SGLang v0.5.21 with HiCache, EAGLE 5/1/6 MTP decode |
| observed price | $16.22 — an hour, observed on 8× RTX PRO 6000; the datacenter classes run $26–$44 |
| catalog id | RadixArk/GLM-5.3-NVFP4 |
The host KV tier is per rank, and that is the trap.
The profile pins --hicache-size 96, and that number sizes each tensor-parallel rank, not the box. At TP8 the cell holds 8 × 96 = 768 GiB of host KV, plus roughly 177 GiB of DSA indexer sidecar — about 945 GiB charged against the box's cgroup memory.max.
The advertised RAM does not gate it, and neither does the 600 GiB filter floor. The only honest read is the cgroup ceiling on the box itself: the reference cell showed memory.max at 1354.93 GiB against 1996 GiB advertised. A box that passes every filter can still fail the tier.
At TP4 the same row holds 4 × 96 = 384 GiB plus the sidecar, which is why the four-card classes are the comfortable shapes and the eight-card class is the one with headroom to spare on compute and almost none on memory.
Download completeness is a launch gate.
The engine's duplicate-safetensors filter aborts if any shard the index names is absent — an interrupted download leaves .incomplete blobs and the loader refuses the model. On a 433 GiB checkpoint at a 1 Gbps floor that is the difference between a paid wait and a paid failure. Let the download finish before the engine starts.
The numbers this shape is known by.
433 GiB the NVFP4 checkpoint 96 GiB of host KV, per rank 1354 GiB the cgroup ceiling a TP8 cell met $16.22 an hour, observed, 8 cards 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/GLM-5.3-NVFP4
or ask the catalog first, with no key and no billing
./target/debug/onefrontier fit RadixArk/GLM-5.3-NVFP4 --tp 8
./target/debug/onefrontier best RadixArk/GLM-5.3-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.