sys1
Canonical version: sys1.
sys1 is an open-source inference server written in Rust by Alvaro Bartolome. It serves open-weight System One models behind an API that copies TypeSafe AI's: you send the same request you'd send to Jev (a state plus typed questions) and get typed answers with probabilities back. Except it runs on YOUR machine.
Why does that matter? Because the main objection to Jev is that it's a closed API you have to send your data to. sys1 keeps the interface and swaps the model for an open one, so code written against Jev can point at a local endpoint instead.
What it does:
- TypeSafe-compatible endpoint.
POST /v1/systemone(alias/v1/decide), with the same three question types:choice,scoreandnoul(see System One Primitives) - Open models only. Today that means Laya:
convaiinnovations/laya(English, guardrails, email triage),laya-multilingual(100+ languages, ~2.2x faster) andlaya-typed-decisions - Built for speed. Tokio, axum and serde for the server, Hugging Face's candle for inference, dynamic token-based batching, SDPA on CPU, Metal and CUDA, and Flash Attention 2/3 on Ampere, Ada Lovelace and Hopper GPUs
- Easy install.
cargo install sys1 --features cpu(ormetal,cuda), thensys1 --model-id convaiinnovations/laya --dtype auto
The example in the README is a good one: a "charged twice, refund me today" support message, with a Choice for the route (billing, bug, account), a Score for urgency on a four-level rubric, and a Noul for "escalate to a human?". Three decisions, one request, no text generated.
Keep in mind that sys1 is a server, not a model. Its answers are only as good as the open model behind it, and Laya isn't Jev: it hasn't been trained with RLCD, so don't assume its probabilities are calibrated the same way. Measure on your own data first.
I like it because it makes the TypeSafe API shape a de facto standard, like the OpenAI chat API before it. Once several servers speak the same format, you can prototype on hosted Jev and move sensitive workloads to a local box without rewriting anything. OpenClaw already took that route with a provider-neutral decision model setting.
References
Related
About Sébastien
Ready to get to the next level?
Found this valuable? Share it with someone who needs it.