browseML

bankML in your browser · verified · a receipt on every answer
the full bankML Space
who answers bankML · free CPU · not bankML's engine · no receipt ↗ your own bankML ↗ a Hugging Face provider · not bankML · no receipt · your quota ↗ ↗ answers on the full bankML Space (PYTHAI/bankml), not on this page

Runs in this tab, on your CPU: bankML compiled to WebAssembly, with Bonsai-1.7B Q1_0 (248 MB, downloaded once, checked against its sha256 pin). A receipt with every answer, checked here. Free. Private: your questions never leave this computer.

model runs on this tab: no server does the work (this Space only serves files) this device reading…
The model is not loaded. Your first question offers the one-time download.
session 0 in · 0 out · 0 turns
An answer cannot be stopped midway: it runs to its end or to this length.
what this is: three builds, the oracle, the limits

browseML is bankML's engine compiled to WebAssembly. Before it answers, it checks the model as bankml serve does: the GGUF guard, then the sha256 of all 248 MB against its pin. Every answer it gives carries serve's receipt.

buildused when
browseml-mt-relaxed.wasmthreads, relaxed SIMD compiled here and fused on this machine
browseml-mt.wasmthreads, otherwise (Firefox, or not fused)
browseml.wasmone thread: the page is not cross-origin isolated (or Safari)

All three are exact, and they give the same answers. On llama-server b11192's nine recorded conversations, every text, finish reason and token count is identical: browseml.wasm 9 of 9 and browseml-mt-relaxed.wasm on 2 threads 9 of 9.

Limits: threads need cross-origin isolation, so Safari runs on one thread. The first download is 248 MB. A tab holds about twice the model while it loads. An answer cannot be stopped midway.

engine log