Glint-Research/glint-2-effort-explorer

Glint-2 Effort Explorer

An interactive demo for Glint-Research/Glint-2 — a 1.06-million-parameter pure-loop transformer: one shared block looped exactly 8 times, with tied embeddings over a 4,096-token BPE vocabulary.

The model is tiny (17 MB) and runs on CPU. This Space faithfully ports the architecture and the six-level effort ladder from the model repo's generate.py and effort.py reference scripts, and loads the weights, tokenizer, and 3.5 KB corrective probe directly from the Hub.

The one rule: loops=8

The model is trained for exactly 8 loops of its shared block. The checkpoint config lists 16 (that is the table capacity and a trap) — running it at 16 produces gibberish. loops is pinned to 8 in this demo.

Effort levels

The levels scale search compute, not the model. loops stays 8 at every level:

From high upward, every rerank decision blends 2.0 * P(real) from the corrective probe (effort_probe.pt) — a one-layer net trained to tell real web text from the model's own generations — so the search steers toward prose rather than confident boilerplate.

Try the same prompt at low and ultra to see effort scaling in action: the ceiling does not move, but the model reaches it far more often.