Each capability, as its own page.
Not a wall of claims — one researched page per technique, in a consistent style: what the problem is, how veizik handles it at a reader’s level, the figures we measured, and a reproduction guide so you can take the number yourself. This set is the basis for the engineering blog.
Video · memory
Video that doesn’t run out of memory
Why peak VRAM stays under a 24 GB card through a 720p WAN render — bounded working set, only the active experts resident, no prompt-driven spike.
Read · under 24 GB →Video · reliability
Resume a render, exactly
A render cut off five times — twice by a host crash — resumes from its last step and finishes to the identical latent hash.
Read · same SHA-256 →Language · single GPU
A 35B MoE on one consumer card
A 35-billion mixture-of-experts model resident on a single 24 GB card in about 21 GiB — no second GPU, no CPU offload — decoding at 194 tok/s on an RTX 3090.
Read · 21 GiB on 24 GB →Language · long context
Decode that holds at a long prompt
Decode stays near-flat out to 16K tokens on a consumer card, and past 128K on a datacenter one — a measured rate at each length, not a spec-sheet maximum.
Read · 128K tokens →Image · low memory
A 12B image model on a small card
An image model that fits a small card by trading memory for speed, with per-architecture GPU code that needs no compile step at render time. Image generation is in preparation; a measured page follows when it is offered.
Coming →System · independence
Its own stack, top to bottom
Each model runs from its own sealed, model-specific package — a single engine binary with no Python, no framework and no numeric library. You can check that yourself, in three commands, on the file you download.
Read · verify it yourself →Pages are published as each is researched and measured. Every figure is veizik’s own; no competitor is named, and no implementation detail is given — only what a reader can hold and reproduce.