Laya: an open 421M decision model, calibrated and CPU-fast
Now topping Hugging Face: a free, self-hostable classifier that returns calibrated typed answers in ~33ms — plus October's model retirements.

Copy markdown
A decision model, not a chatbot
Laya skips token generation: one ~33ms forward pass turns text, an email, or JSON into a typed answer — a choice, score, or yes/no — with a calibrated probability attached. It's Apache-2.0, installs with `pip install laya`, and at 421M (English) / 322M (multilingual, 100+ languages) it runs on CPU, so you can drop a self-hosted router, intent classifier, or content gate into an agent with no per-call API bill.
Faster and better-calibrated than the paid System-1 models
On its own card Laya claims ~7.8× faster inference than TypeSafe's Jev decision model and tighter calibration (ECE 0.081 vs 0.144 after temperature fitting), with a fine-tuned variant at 0.766 on the typed-decisions benchmark. Calibration is the point: the confidence it returns is one you can actually threshold on in an agent's control flow instead of guessing.
Model-retirement watch: what breaks this month
OpenAI's deprecation page lists gpt-5.4-cyber as retired today (Oct 1) and sunsets its legacy GPT snapshots on Oct 23, remapping them to the gpt-5.6 terra/sol/luna line — so any hardcoded model IDs in your automation need updating now. GPT-5.5 is also reportedly scheduled to shut down Oct 14 across ChatGPT, Work, and Codex.