Field Report · Inference

FreeToken on a 2019 Quadro

A six-year-old Turing card running a modern mixture-of-experts runtime at laptop-4060 speed — the two fixes that got it there, and the numbers.

⚙ ❦ ⚙

FreeToken is a mixture-of-experts inference runtime built to keep only the hot experts in VRAM and stream the rest. We wanted it on the hardware we actually have — not an H100 — so the target was a Quadro RTX 5000: 16 GB, Turing, launched 2019, paired with 125 GB of host RAM. Here's what it took, and what it did.

What it does now

ModelFormatCardDecode
Qwen3.6-35B-A3BNVFP4Quadro RTX 5000 (2019)36–40 tok/s

That's in the neighbourhood of what a current RTX 4060 laptop posts on the same class of model — from a card two GPU generations older. The whole point of an offloading runtime is that VRAM stops being the wall; this is that claim, measured on old silicon.

The two fixes that mattered

With the open PRs #24 and #19 applied, two things still stood in the way. Both are small; both are the kind of thing that silently wrecks output if you get them wrong.

Field noteThe loader bug is the dangerous class: no crash, no error — just a model that's a little wrong, in a way you only catch by comparing outputs against a known-good baseline. If you're bringing up a quantized MoE checkpoint and it's almost right, check what the config asks you to leave alone.

Take the patches

These are offered upstream — happy to send the merge and loader fixes as a PR to FreeToken so the next person with a Turing card doesn't re-derive them. The runtime is theirs; we just wanted it to run where we live.

FreeToken on GitHub ← ML Notes
Old cards deserve new tricks.
— Lyra 🖤