Decis

The README publishes no quality comparison against hosted Jev, because Decis reimplements the interface and not the model. Backbones, weight sizes and per-engine limits are in docs/engines.md, and docs/performance.md carries measured latency and memory on the machine they were measured on. It also ships its own self-audit in docs/design-review.md listing defects found and what is unverified.

The wire contract is versioned v1 and documented as stable, only ever gaining fields, and the server reports its own version at /healthz. Beyond that the README is candid that this is pre-1.0 at v0.4.0, and the repository carries a design self-audit describing defects found, methodology limits and what remains unverified, which is worth reading before depending on it.

Not with Docker. The published multi-arch images carry their weights inside, so nothing is downloaded at startup and there is no volume to mount; the engine id is the image tag. From a source checkout it is different: decis download --engine laya-multilingual pulls 647 MiB once, and you need uv sync --all-extras or a named engine extra, since a plain uv sync installs none and removes what an earlier sync added.

No. The server requires a bearer token, compares it in constant time, and refuses to start on a public address with no token configured. The quickstart still expects you to set DECIS_API_KEY in .env, and the playground keeps its key server-side so the browser pages never see it. Without a key, jev_config-style local use is not the path here; the token is mandatory.

Six ship. laya-multilingual is the default and covers about 100 languages; kev-0.8b, jeff-qwen3.5-0.8b and jeff-gemma4-e2b have published multi-arch images, while laya and laya-typed-decisions run from a source checkout. One image per engine means the engine id is the tag. Backbones, weight sizes and per-engine limits are in docs/engines.md.

One changed line. Point the official SDK at your own host with TypeSafeClient(api_key="local", base_url="http://127.0.0.1:8000"). You do not pass model=, because the SDK sends its jev-latest default and Decis answers with the engine it is actually running. The examples directory holds runnable curl and official-SDK examples that CI executes.

Two probes. /healthz answers immediately, while /readyz reports loading until the model can actually respond, so the quickstart polls readyz in a loop before issuing any call. For diagnosis, decis doctor prints the device, the torch build and every GPU the driver reports, and decis models lists what the server is serving.