You can read the engine.
Most inference endpoints ask you to trust a black box. This one runs on memra, an open-source engine in Rust and CUDA — the same code that serves your requests is the code you can clone, read and disagree with. That is the whole credibility argument, and it is checkable in about a minute.
- Who
- Avi Fenesh — Engineer, tiyuvta. I wrote the engine, I run the hardware, and I answer support@tiyuvta.ai. There is no support tier and no queue between you and the person who built it.
- What it runs on
- memra: An open-source inference engine in Rust and CUDA. It is the engine this API runs on, so the serving path behind the endpoint can be read rather than taken on trust. Rust and CUDA · licence · github.com/avifenesh/memra
- The lab
- The research and the receipts live at tiyuvta.ai. This endpoint is how that work pays for the hardware it needs.
What that buys you, concretely
The quantisation is checked, not asserted
Serving runs NVFP4 4-bit weights. Every build is verified against an eager reference and the vendor's higher-precision release before it takes traffic, and speculative decoding is proven hash-identical to plain decoding. The gate list names each check.
Numbers arrive with their conditions
138–141 tok/s single stream on build v0.86.1, and the 74.9 tok/s floor next to it. First token 0.166s cold. Run your own during the trial credit and trust those instead — the full table.
The price sheet has no asterisk
$0.38 in, $0.20 cached in, $2.60 out per million tokens, no request fee, no subscription. Cached input is billed when it happens and we publish no hit rate, because we cannot promise one. The sheet.
What I will not claim
No uptime promise
There is no SLA. A sustained run of 576 requests with zero errors and zero shed is what exists; a multi-month availability record does not yet. Do not make this a single point of failure while that is true.
No capacity theatre
You will not find a tokens-per-second-per-fleet figure here. Capacity is scaled to demand, so publishing a ceiling would be either a boast or a limit — and neither is a fact you can plan against.
Small on purpose, and it shows
One model live, one more in bring-up, one person answering support. The upside is that nothing between you and the serving path is hidden. The downside is real and this is it.
Free credit is trainable
Prompts sent on the 10-dollar trial credit may improve our models. Paid traffic is not used for training unless you switch it on for a rebate. Both halves are in the privacy terms.
$10 of credit, one email, no card.
Read the engine first if you would rather — github.com/avifenesh/memra. It will still be there afterwards.