Mosaic 1

A robust no-limit hold'em model for 2 to 6 players.

Model size
26M parameters
150+ decisions / second / CPU core
Speed
150+
Decisions per second, one CPU core
Vs Slumbot
+12 bb/100
(+12.6 ± 5.7 bb/100)
Opponent adaptation
up to +711
bb/100 vs fixed archetypes, ablation-tested

A different kind of poker AI

No CFR blueprint or sub-game solvers. Competitive against fixed GTO bots, but designed for robust play against humans with arbitrary action spaces and stack sizes. Mosaic 1 is a small neural agent that plays the actual game — full action space, real multi-street dynamics, six-handed tables — and adapts to the opponents at the table as hands accumulate. To our knowledge the 6-max model is the first publicly available agent of its kind.

We publish our evaluation methodology with the model: solver-anchored audits, luck-adjusted archetype tables, and exploitability probes that are stricter than the published literature protocols. No superhuman claims — measured ones.

Play it from four lines of code

curl -X POST https://api.icosalab.com/v1/tables \
  -H "X-Api-Key: $KEY" -d '{"game": "6max"}'
{"table_id": "…", "hole": ["Ah","Kd"], "legal": [
  {"slot": 0, "kind": "fold"}, {"slot": 1, "kind": "call"},
  {"slot": 4, "kind": "raise_to", "to_chips": 1500.0}], …}

The server deals; your bot plays any legal action. One persistent table per session — the model's read on your play warms up across hands, exactly as it would against a human. Full protocol in the docs.

The research tier is free

Confirmed email, generous per-key rate limits, both games. We log hands for research and training — that's the deal, and it's in the terms. Real-money play and use against real-money platforms are prohibited.