Tiny text, published with Loom
example/tiny-text is deterministic arithmetic, not a trained chatbot. It
accepts a, b, and x, uses # as EOS, and has one four-byte checkpoint
parameter. Prompt abxa with max_tokens: 2 produces bx. The checkpoint is
small enough to embed in the model card; fetch materializes it as safetensors.
Execution uses real JIT-compiled kernels, pooled state, and the text service.
Token IDs are a=0, b=1, x=2, and #=3. Starting at zero, each token updates
h = (5*h + token + 2) % 97 and k = (7*k + h + position) % 101, then appends
k to history and advances position. Prediction is (h + sum(history)) % 3,
or EOS every eleven consumed tokens. This is a small retained-state example,
not a performance proxy. Tools, reasoning, and images are not supported.
With the Loom SDK
installed, python smoke.py checks the catalog and locks, fetches this model,
and checks its output on a private CPU server. python refresh.py explicitly
updates asset locks and the catalog after authoring changes. Both scripts have
--help; neither needs a source checkout or build system.
The Hugging Face publishing guide owns the sharing workflow, upload commands, and examples of external weights. This repository owns its model programs, metadata, and output check.