blink-tiny
A 407K-parameter byte-level decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module. Given a state, a question and options declared at call time, one pass returns a probability per option; scoring allocates no memory.
Not a language model: it is trained per task and learns the form of a decision, not the meaning of text. SQLite AI ships int8 .blink containers in the GitHub repo, not on Hugging Face, so this page links them instead of listing files; they need Blink's own C or WebAssembly runtime. On its own held-out synthetic corpus the maker reports routing 0.998 and judgment 0.999, but comparison and selection stay at chance; the pooled figure depends on the task mix. On WANLI it reaches 0.491 top-1 (chance 0.333). Unrelated to Pixilab's Blink. The WANLI-trained containers derive from CC BY 4.0 data.
What it decides
- choice — picks one option from a set
- route — sends the input to one of several destinations
At a glance
| Parameters | 407K |
| Maker | SQLite AI |
| Released | 2026-09-22 |
| License | apache-2.0 |
| Reported accuracy | 63.3% |
| Reported latency | 406 µs p50 per fresh decision (256-byte state, 4 options) on one Apple M5 Pro core; 54 µs with the state reused |
Read more
This page was opened by System One for SQLite AI, who can claim the organisation and take it over at any time.