kime
Early Rust runtime that answers Jev and Laya typed questions, currently serving Laya checkpoints on CPU, CUDA and Apple GPUs.
Overview
kime (from the Japanese for 'the decision') answers typed questions about text: a state such as an email, ticket, JSON document or web page, plus choice, score and yes/no questions, with a calibrated probability for every possible answer and no generated text. It is written in Rust and speaks both Jev's and Laya's wire formats, so the same request works against Jev's API, Laya clients work unchanged, and Laya's Python API works with an import change.
Its stated goal is to be ten times faster and cheaper than the best baseline (Laya on its own hardware, laya-mlx, laya-coreml and hosted Jev) on every speed and cost benchmark while matching or beating them on accuracy and calibration. The README is clear that this is a target, not a result: today kime runs the published Laya checkpoints on CPU, CUDA and Apple GPUs with the same answers as Laya, and kime serve answers the System One API over HTTP. Native kime models, a router, caches and SDKs are still to come, and the design lives in a spec folder.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should handle this? | billing · technical · sales · other |
| Score | How urgent is it? | not urgent · this week · today · right now |
| Yes / no | The customer threatens to leave | — |
How it works
- kime pull laya to fetch a checkpoint
- kime serve starts the HTTP server
- Send state and typed questions to /v1/systemone
- Receive per-option probabilities in Jev or Laya response shape
Strengths
- Compatible with both Jev and Laya clients
- Runs locally on CPU, CUDA and Apple GPUs
- Ambitious targets published with measurement commitments
Caveats
- Early: native models, router and SDKs not built yet
- Performance targets not yet measured
- Currently depends on Laya checkpoints
Getting started
kime pull laya && kime serve --port 8000, then POST to localhost:8000/v1/systemone.At a glance
- Type
- Project
- Status
- prototype
- Language
- Rust
- Runtime
- CLI and HTTP server
- Has code
- yes
Use cases and integrations
How curated lists describe it
Typed decisions over text in milliseconds.
wh000wh000/awesome-jev-live
Frequently asked questions
Is kime faster than Jev today?
Not shown yet; the README lists speed targets, not measured results.
Which models does it run?
Currently the published Laya checkpoints; native kime models are planned.
Can I use a Jev client?
Yes, it accepts Jev's request format.
Related projects
Docker service exposing local Laya, Von, a CLM head or hosted Jev through one TypeSafe-compatible HTTP and MCP endpoint.
MCP serverbetaPythonConverts open Jev reproductions (Laya, Kev, NanoJev, PlayJev) to INT8 ONNX for offline CPU decisions at about 15.6 ms per question.
ServicebetaPythonNative Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.
ProjectbetaGoGo harness for System One models: hosted Jev, open Kev in-process via llama.cpp, and Cloudflare Clef, with typed helpers and a CLI.
ServicebetaGoGo gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.
ServicebetaGoRuns Intern-Decision-0.8B on Apple Silicon behind a local /v1/systemone endpoint that accepts screenshots for agent decisions.
ServicebetaPython