Chapter 4 of 6

In your browser: download, check and run

The model arrives in parts, and each part is checked before it is kept. A small C engine compiled to WebAssembly then does the arithmetic on your own processor.

Fig. 1Parcel line

Space or Right steps the run

first visit · 0 of 17 parts stored · 0.0 of 534.8 MiB

SHA-256 match, or a part held in the cache
a part of the model

Fig. 1 — Parcel line. Flyingkiwi 0.6B's weights arrive as 17 parts, each checked against its SHA-256 fingerprint before it is kept in the browser's Cache Storage; later visits read the parts from there and check them again, and only a part that fails its check is downloaded again. The pins and ridges are an illustration of that comparison, not a model of the hash itself.

Fig. 2Row stamper, one head or six

Space: 1 or 6 threads · Right: step

6 threads · 16 numbers per instruction · rows done 0 of 32

row being stamped now
stamped so far

Fig. 2 — Row stamper, one head or six. One WebAssembly instruction multiplies and adds 16 eight-bit numbers at once, and up to six threads each take the next rows from a shared counter (the rows drawn are a few of the real ones). On one laptop with six threads, Blazing Kiwi wrote 98–127 tokens per second.

In numbers

Parts of up to 32 MiB

Flyingkiwi 0.6B English: 560,823,360 bytes of weights, three parts at a time.

SHA-256

Checked for every file

Against the size and fingerprint in the manifest, when downloaded and when read back.

Numbers per instruction

One WebAssembly SIMD instruction multiplies and adds sixteen 8-bit numbers.

98–127

Tokens per second, Blazing Kiwi

Six threads on one laptop processor (AMD Ryzen 5 7530U) in Chrome.