models · deep reasoning
DeepSeek-41-Flash, for the hard work.
The reasoning lane: a model tuned for correctness on structured work and long inputs, shipped as a precompiled package with its CUDA kernels and captured graphs baked in. It is not a vLLM row, and the difference matters.
| weights | MXFP4 experts with compressed CSF scales, BF16 attention, FP8 KV; 290 GiB card-side; 513.5 GB on disk including the host-resident tables |
|---|---|
| context | 1,000,000 tokens |
| hardware | 4× RTX PRO 6000 (Blackwell SM120, 96 GB each) |
| tensor parallel | 4 — the package is fixed-shape TP4 |
| engine | Flare 0.1.0 (built 2026-10-19): one Rust process, no Python, no runtime compilation |
| observed price | not measured — under the same $10.00/h cap as the flagship class; no capture has priced this cell |
| catalog id | local-inference-lab/DeepSeek-V4.1-Flash-1017-FLARE |
The package is the runtime.
The 1017 re-release is a signed Flare package: weights, compiled CUDA kernels, and captured graphs in one fixed shape. vLLM, SGLang, Transformers, and llama.cpp cannot load the format at all — the engine is a single 20 MB Rust binary, and the pack is prepared for exactly four 96 GB SM120 cards.
The card-side of the pack is ~290 GiB, so 72.5 GiB per card under the 80% headroom the profile pins. The remaining 188.8 GiB is Engram tables that ride host RAM in the recommended mode, which is why the host floor is 300 GiB and the disk floor is 800 GB — disk is fixed at create and the nine runtime files alone are 513.5 GB.
On the card, that leaves a 21.16-million-token KV pool. The concurrency shape is captured rather than tuned: max-concurrency 32, five-draft adaptive packed MTP verification, the throughput scheduler.
Two failures the kit encodes so you do not repeat them.
The vendor's flare image is debian-slim plus the binary, with no openssh-server — vast's ssh launch mode dies at its own sshd step and the onstart never runs. A live cell sat with idle GPUs and zero bytes written before that was understood, so the launch vehicle is an sshd-carrying carrier image and the pinned flare binary is fetched from the registry at boot. Second: the publisher signing-key prompt is interactive by default, so an unattended first boot needs the publisher namespace pre-accepted and a writable trust store, or it waits on a prompt that has no tty.
The numbers this shape is known by.
290 GiB card-side weights at TP4 513 GB on disk, nine files 21.16M tokens of KV pool 0 Python processes at runtime 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/DeepSeek-V4.1-Flash-1017-FLARE
or ask the catalog first, with no key and no billing
./target/debug/onefrontier fit local-inference-lab/DeepSeek-V4.1-Flash-1017-FLARE --tp 4
./target/debug/onefrontier best local-inference-lab/DeepSeek-V4.1-Flash-1017-FLARE
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.