tiyuvta — conclusive refutation by evidence

tiyuvta is an AI lab where the evidence decides.

tiyuvta studies model efficiency end to end — draft-head training, expert pruning and healing, precision assignment, speculative decoding — and measures every claim in memra, an inference engine built from scratch to expose what a number includes. The result can be yes, no, or not yet; the failed arms stay visible.

Read the research See the inference product

170.5 tok/s single-stream, spec decode, through the serve surface — c=1, N=5 median · RTX PRO 6000 Blackwell 96GB (rented pod)
0.182 s / 0.003 s TTFT cold / warm — cold forced with a fresh cache_salt per request · RTX PRO 6000 Blackwell 96GB (rented pod)
same tokens, alone or batched greedy output byte-identical at c=1 vs c=16 — same server, 16/16 on four models

fig. 01 — the apparatus is part of the argument

memra — the engine

A from-scratch Rust+CUDA inference engine, public and MIT-licensed: MTP speculative decoding, FP8 checkpoint serving, MoE spill, exactness gates on Blackwell. Every kernel bit-audited against a CPU reference; every published number's raw run logs committed in-repo.

github.com/avifenesh/memra

Research — trained, not only measured

Compact MTP draft heads with their own vocabularies, pruned-and-healed MoEs, precision allocation under byte ceilings, training-recipe studies. Preregistered kill gates; refuted hypotheses published with the same care as supported ones.

the evidence ledger →

Inference — the product

The lab serves the fastest deterministic inference its engine can produce, as an OpenAI-compatible API. Serving pays for the hardware the research runs on. One model class, commitments you can verify from your own client.

the landing page →

fig. 02 — recent entries in the ledger

2026-08-05
Cold TTFT: llama.cpp wins the head-to-head, 0.19 s vs 0.53 s bounded limit
2026-08-05
The train-to-serve loop closes: 32 minutes, every gate green win
2026-08-04
27B spec decode through the serve surface win
2026-08-04
nvcc 13.0.88 miscompile caught by the kernel bit-parity gate win

Full ledger — wins and bounded limits both →

fig. 03 — writing

Serving is a determinism problem 16% of (layer,token) pairs
The compiler ate my byte 1 byte store, dropped at -O3
Dogfooding found three bugs in my server 3 bugs, 1 class

All posts →