Sezika

Paused C#/.NET 10 non-autoregressive decision engine running a Laya/mmBERT checkpoint locally, returning Choice, Score and Boolean answers.

ModelresearchApache-2.0

View on GitHub ↗

Overview

Sezika is a local multilingual System 1 decision engine written in C#, documented mainly in Chinese. It takes text or JSON state and returns typed judgments with probability distributions: Choice (selected option and full distribution), Score (probability-weighted score and level distribution) and Boolean (P(true)). The model, CPU and CUDA kernels and scheduling are implemented in C#, with CUDA via build-time PTX and static driver bindings, so core inference does not depend on PyTorch, ONNX Runtime, cuBLAS or remote APIs; Native AOT has been validated on Windows and Linux WSL2. It currently runs a pinned Laya/mmBERT checkpoint.

The project is explicitly paused pending compute and funding. Full-parameter training was started on an RTX 5090 but only about 0.31% of the planned 14.6 million updates were completed before the server was shut down, and the README estimates roughly 40 days of GPU time to finish. Outputs are uncalibrated and multilingual quality, calibration and sealed testing remain to be done. Code is Apache-2.0.

What the decision model decides

Type	Question	Answers
Choice	Which candidate best matches the intent?	—
Score	Score the state against ordered criteria	—
Yes / no	Is this proposition true? (Boolean)	—

How it works

Prepare and verify model assets
Run the CLI predict command with a JSON decision request
Engine encodes state and scores options in C# CPU/CUDA kernels
Caller applies business logic to the typed answer

Strengths

Pure C#/.NET stack with Native AOT
No Python or ONNX dependency for inference
Local, offline operation

Caveats

Project paused; training about 0.31% complete
Probabilities uncalibrated
Docs mostly in Chinese

Getting started

dotnet build Sezika.slnx -c Release, then dotnet run --project src/Sezika.Cli -- predict --model .artifacts/models/laya-mmbert --input samples/requests/decision.zh.json

At a glance

Type: Model
Status: research
Language: C#
Runtime: .NET 10 CLI, CPU/CUDA
Has code: yes
License: Apache-2.0

Use cases and integrations

Infra / SDKs / Integrations537

How curated lists describe it

Local inference: an independent, Jev-inspired non-autoregressive C#/.NET 10 engine that runs a pinned Laya/mmBERT checkpoint locally through C# CPU/CUDA kernels, with documented Native AOT validation, and returns its own typed `Choice`, `Score`, and `Boolean` answers to .NET applications; probabilities remain uncalibrated and multilingual quality is still under evaluation.
yibie/awesome-jev

Frequently asked questions

Is Sezika usable now?

The engine runs, but training is paused and outputs are uncalibrated.

Does it need Python?

No; inference is implemented in C# with its own CUDA path.

Which model does it run?

A pinned Laya/mmBERT checkpoint for now.

Related projects

decidealot
Docker service exposing local Laya, Von, a CLM head or hosted Jev through one TypeSafe-compatible HTTP and MCP endpoint.
MCP serverbetaPython
edgejev
Converts open Jev reproductions (Laya, Kev, NanoJev, PlayJev) to INT8 ONNX for offline CPU decisions at about 15.6 ms per question.
ServicebetaPython
go-system-one
Native Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.
ProjectbetaGo
gojev
Go harness for System One models: hosted Jev, open Kev in-process via llama.cpp, and Cloudflare Clef, with typed helpers and a CLI.
ServicebetaGo
hearim
Go gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.
ServicebetaGo
intern-decision-mlx
Runs Intern-Decision-0.8B on Apple Silicon behind a local /v1/systemone endpoint that accepts screenshots for agent decisions.
ServicebetaPython