Precision
Fewer bytes per weight, without a converted copy.
The sub-byte numeric formats are proprietary — what is shown here is the shape of the idea and the output it produces, not how any format works.
Precision · numeric formats
Fewer bytes per weight, without a converted copy.AI-generated · Stable Diffusion 3.5 · RTX 3090 · Powered by Stability AI · drawn at sub-byte precision
1artifact · every mode, no file per precision
0conversion steps · precision injected at load
per layera format chosen per layer & architecture
One checkpoint, every precision
Read the weights you already have; the precision is a setting.
No converted copy, no file per bit-width.
- One original checkpoint, every precision. veizik reads the published safetensors you already have and chooses a numeric precision in the engine, as it loads. There is no quantized copy to produce first — no per-bit-width file to download, no calibrated checkpoint, no second artifact to keep in step. The precision is a setting, not a file.
- Switching precision never re-packs the weights. Because the format is applied in-engine, moving between a memory-lean and a quality-lean setting is a flag — no re-export, no second download, no re-upload to the GPU.
- A capability veizik shares, and then extends. Reading an original checkpoint and applying precision at load is something a handful of engines now do. What is less usual is carrying several execution modes in one artifact and switching between them without rebuilding the weights — and the breadth of the format catalogue and the per-layer mix below.
One artifact, every mode
Several execution modes from one stored representation.
Switch quality for speed with a flag — no re-pack, no second download.
- The headline is not a smaller file. Compact weights are table stakes; a strong low-bit build is a known quantity. What one stored representation buys is several execution modes from the same bytes — quality, balanced, fast — chosen at run time.
- Switch the mode, keep the artifact. Moving from the highest-fidelity mode to the lowest-cost one is a command-line flag. There is no second model to download, nothing re-packed on disk, and nothing re-sent to the GPU — the stored representation does not change.
- The idea is not new; the integration is the point. Running one model at more than one precision has prior art in the literature. What veizik brings together is the storage, the load, the precision switch and the per-stage schedule inside a single native engine — across language, image and video models, from a datacenter GPU down to a Raspberry Pi.
A catalogue, not a menu
A format per layer, from a growing ladder.
Many sub-byte formats, mixed within one model and injected with no conversion step.
- A catalogue, not a menu. Most engines expose a short list of widths — commonly 1.5 to 8 bits. veizik draws from a much larger, still-growing catalogue of its own sub-byte numeric formats and mixes them within one model: a bandwidth-bound layer and a precision-sensitive one need not share a format.
- Injected at compile time — zero conversion. The precision is lowered into the engine’s own kernels, so there is no separate dequantization pass in the hot path.
- Designed to grow. New numeric formats are added as the research lands — the newest are still being wired in — without a new kernel or a re-pack. The ladder is meant to get finer over time, not to freeze at a number.
Quality per bit
Fewer bits only count if the output holds.
The output is the evidence; the figures per device are linked.
- The honest measure is quality per bit. Fewer bits only help if the output holds, so the number that matters is the effective bits-per-weight at a quality bar — counting the real stored size, not a nominal label. veizik tunes a format to the quality it has to keep, then to the bytes.
- Judge the output, not an adjective. Every image on this page — and the renders in the gallery — was drawn by the engine at sub-byte precision; the figures per device and stage are on the edge and benchmarks pages.
- Not a universal claim. A frontier engine specialised for one model may reach it at fewer bits; the breadth of the catalogue, the per-layer mix and the single-artifact mode switch are what veizik brings across models, architectures and hardware.
Stable Diffusion 3.5 — Powered by Stability AI; FLUX.1-schnell — Black Forest Labs (Apache-2.0); SDXS on a Raspberry Pi 5. Every image is veizik’s own render, drawn from the original checkpoint at sub-byte precision.