models · the fast flagship
GLM-5.3-Flash, on four cards.
An open-weight flagship model quantized to fit four workstation GPUs. It is the kit's default launch profile — the shape the whole path was proven on, fit gate through tunnel.
| weights | NVFP4 (modelopt); 184 GiB |
|---|---|
| context | 1,048,576 tokens, native |
| hardware | 4× RTX PRO 6000 (Blackwell SM120, 96 GB each) |
| tensor parallel | 4 |
| engine | vLLM, fp8_ds_mla KV cache, DFlash2 speculative decode at depth 7 |
| observed price | $5.49 — an hour, observed on the live four-card cell; the class band is $4.70–$10.00 |
| catalog id | local-inference-lab/GLM-5.3-Flash-NVFP4 |
Why this model is the default.
It is the one row in the catalog where a capable open model, an ordinary market segment, and a budget that does not require a company all line up at once. The four-card RTX PRO 6000 class rents at single-digit dollars an hour in the EU, and at tensor-parallel 4 the checkpoint is 46 GiB per card — a third of the ceiling left for the KV pool and the CUDA graphs.
It is also the row the kit was proven on first. Every stage of the path — filter, fit gate, create, engine boot, tunnel arm, verify — ran on this profile before any other. That is why it ships as the default rather than as one option among equals.
A speculative draft model rides beside it: local-inference-lab/GLM-5.3-Flash-DFlash2 at depth 7. It is a second gated checkpoint, and the launch profile names both, so a run does not silently lose the decode speedup.
The TP2 trap.
Two cards look like enough and are not. At tensor-parallel 2 the weights are 92 GiB per card; the load runs to about 87% and dies on a 96 GiB card — on the paid clock, because billing starts at loading. This capture is the fit gate's origin. The only 5.3-Flash weight that genuinely fits TP2 is a 62 GB expert-pruned pack, and the public one is compressed-tensors, which mismatches the modelopt quant the serve script pins — a TP2 cell needs the gated pruned twin and a quant override, not merely a smaller model path.
The numbers this shape is known by.
46 GiB weights per card at TP4 87% where the TP2 load died 1M token context, native $5.49 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 local-inference-lab/GLM-5.3-Flash-NVFP4
or ask the catalog first, with no key and no billing
./target/debug/onefrontier fit local-inference-lab/GLM-5.3-Flash-NVFP4 --tp 4
./target/debug/onefrontier best local-inference-lab/GLM-5.3-Flash-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.