Support the work
Open infrastructure, funded in the open.
Runner is developed openly rather than around a subscription. The engine is free forever under Apache 2.0, and that is not a launch promise but a published rule: no priced engine feature, no claim weakened to sell something, no license change. What keeps the project moving is measurement time on real hardware, and that is what a contribution buys.
One-off contributions through Stripe. No tiers, no perks, no gated features; the receipts of what it bought are the evidence pages.
What contributions enable
Every table on the evidence page cost hardware time.
The project's distinguishing asset is that its numbers hold up, including the negative ones. Numbers that hold up come from running the same measurement on more than one machine, against the other engines, on the real files. That is the budget.
Measurement hardware
The CUDA tables were measured on a Blackwell MIG slice and an RTX 3070; the Metal work on an 8 GB M1 and, for a two-day window, a loaned 128 GB M5 Max that produced the Metal dispatch work, the 120B idle-coexistence measurement and the big-model certification sweep. Loaner windows end. Owned hardware means a measurement can be repeated when a reader asks.
Cross-engine validation
The truncation benchmark required standing up vLLM, llama.cpp, Ollama, TensorRT-LLM and SGLang on the same box, at pinned versions, and re-running the ladder. Every compatibility row pins a llama.cpp revision to compare greedy output against. That work is where most of the hours go, and it is the work that makes a comparison mean something.
Model hosting and artifacts
Sixteen repositories on Hugging Face, several of them 17 to 18 GB files built, gated and re-measured when the bar changed. Publishing a derivative with its envelope, and keeping a near-miss published with its numbers, has a storage and bandwidth cost that a marketing page does not.
Independent reproduction
Runner credits external reproductions in the docs, and the first one, on a Tesla T4 the project never touched, sharpened the determinism claim. Funding the hardware to reproduce a reader's finding, success or failure, is how the claim boundary stays honest.
How the project is funded
Three sources, none of them the engine.
Zenova AB takes engagements around local inference: private fine-tuning for data that cannot leave the building, with the receipts as the deliverable. This funds the hardware. The engine is never the thing being sold.
One-off, through the Stripe link above. They go to measurement time, hosting and hardware, as described on this page.
A benchmark on hardware the project does not own, an unusual GGUF that fails to load, a coding agent that does not complete its loop, an independent run of the determinism claims. Open an issue with runner --version, runner --caps, the exact filename and the load log. Reproductions are credited.
Analytics on this site, a cookie banner, a paid tier of the engine, or a claim the repository does not make on the day the site is built.