Organization repository on GitHub
Multilingual, non-autoregressive System 1 decision engine in C# and .NET 10. Native AOT and C# GPU kernels via device drivers.
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Models & Reimplementations
Sezika
An independent, Jev-inspired System One engine written in C# and .NET 10: a pinned Laya/mmBERT checkpoint runs through pure C# CPU and CUDA kernels with Native AOT, returning typed Choice, Score and Boolean decisions to .NET applications. Probabilities are explicitly uncalibrated.
- Stars
- 5
- Language
- C#
- License
- APACHE-2.0
- Published by
- Community
- Last verified
- Oct 2026
Used inJev for Software & DevTools
Highlights
- Non-autoregressive engine: a multilingual encoder plus a decision head returns Choice, Score and Boolean answers without generative decoding.
- CPU kernels are pure C# scalar/SIMD; GPU runs through C# kernels exported to PTX by ILGPU and driven via the CUDA Driver, with no cuBLAS or cuDNN.
- Native AOT validated on Windows win-x64 and Ubuntu WSL2 linux-x64 across four backends with 1/8/32-question fixtures.
- A quality retest compared 574/574 PAWS and Nimble rows against a reference; 68.00% and 42.28% were answered correctly in that round.
- Apache-2.0; the README states it does not reproduce Jev's closed model or inherit its speed or quality claims.
Watch out
Apache-2.0, but model weights are not released with the repository and decision probabilities are explicitly uncalibrated; the README says per-language quality acceptance is still pending.
Questions about this entry
Short answers drawn from the linked project or page.
No. You prepare and verify a fixed Laya/mmBERT model package as described in the standalone-usage doc, then load it with the CLI or DecisionModelRuntime; the repository ships the engine, tokenizer, kernels and tooling.
No. It is an independent, Jev-inspired engine built on an mmBERT-class encoder. The project states it does not reproduce Jev's closed model or inherit upstream speed or quality conclusions, and its own outputs are uncalibrated.
Build with dotnet build Sezika.slnx -c Release, then use the CLI: inspect checks a fixed model package, and predict takes --model plus --input (samples/requests/decision.zh.json and decision.en.json are editable) and a --deadline-seconds budget, returning Choice, Score and Boolean answers as JSON. A C# host instead calls DecisionModelRuntime.Load(...) and Evaluate(...) to reuse one session.
Choice selects from candidates supplied in the request and returns the pick, the full probability distribution and its concentration. Score scores against an ordered rubric and returns a probability-weighted expectation plus per-level probabilities and the level description. Boolean returns P(true), which a compatibility layer can map to noul. The engine advises only: the caller controls action and permissions, and open-ended replies belong to a generative model.
Inference operators are implemented in C#, with no Python, PyTorch, ONNX Runtime, cuBLAS, cuDNN or second service process. GPU execution goes through the NVIDIA CUDA Driver: ILGPU 1.5.3 exports the C# kernels to PTX and ABI at build time, and Native AOT invokes them through static driver bindings. The JSON layer is source-generated so AOT and trimming stay analysable.
Requests that would need truncation are rejected by default; an explicit length_policy of laya_compatible switches to compatible truncation and returns diagnostics instead. A token-coverage retest put PAWS at 250/250 constructible in both modes and Nimble at 306/324 strict versus 324/324 compatible, but the project is clear that those are input constructibility rates, not inference success or answer accuracy.
Native AOT was validated on Windows win-x64 and Ubuntu WSL2 linux-x64 across four backends with 1/8/32-question fixtures, and the README warns that a WSL2 smoke test does not stand for bare-metal Linux performance. On .NET 10.0.11 with an i9-13900HX and RTX 4070 Laptop, CUDA long-32 measured 38.319 s end-to-end for a single sample and SIMD long-1 42.183 s; the CPU long-32 cell was never started because it blew the tool's 30-minute cap. The W8A32 quantized path is slower than SIMD and does not reduce total resident memory.
More Jev resources like this
1.3kGitHub stars
A local daemon that pulls open decision models and serves them on TypeSafe's /v1/systemone, so a Jev client switches hosts with TYPESAFE_BASE_URL.
Jev Models & Reimplementations#community#rust#open-models
Community
ollaya-dev/ollayaRust
ollaya-dev/ollayaRust10GitHub stars
A local Jev-compatible server that reads Choice, Score and Noul answers from the logits of any MLX checkpoint or OpenAI-compatible endpoint with no training, serving POST /v1/systemone as a drop-in for the official TypeSafe SDK.
Jev Models & Reimplementations#community#mlx#calibration
Community
bradAGI/rulingPython
bradAGI/rulingPython3GitHub stars
A training-free runtime that turns dense Qwen3 or Qwen3.5 checkpoints into a Jev-style Noul, Choice and Score service, sharing one state prefill across isolated questions and reading each answer from candidate-token logits.
Jev Models & Reimplementations#community#qwen#local-inference
Community
HopLee6/QwevPython
HopLee6/QwevPython877GitHub stars
A local server that reads next-token probabilities from a fine-tuned Spark model and returns Choice, Score, and Noul answers with zero generated tokens.
Jev Models & Reimplementations#community#python#open-models
Community
Rizzo-AI-Academy/rizzo-flowPython
Rizzo-AI-Academy/rizzo-flowPython130GitHub stars
An open recipe for a 27B multimodal decision model trained with full-weight SFT on 73,000 examples over one H200, serving choice, noul and score on /v1/systemone with per-checkpoint provenance and released calibration plots.
Jev Models & Reimplementations#community#calibration#qwen
Community
denis-pplx/autojevPython
denis-pplx/autojevPython77GitHub stars
An Apache-2.0 0.8B decision model that returns calibrated probabilities for yes/no, choice and score questions on CPU. Distributed as GGUF for llama.cpp and served behind a Jev-style /v1/systemone API, so existing Jev clients mostly need only a new base URL.
Jev Models & Reimplementations#community#gguf#local-inference
Community
kouhxp/gutsyHTML
kouhxp/gutsyHTMLFrom the Jev guides
Original write-ups on what this entry is for.
Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.
Jev Models & Reimplementations#multimodal#open-models#decision-model
Read article
One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.
Jev Models & Reimplementations#decision-model#benchmarks#pricing
Read article
Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.
Jev Models & Reimplementations#open-models#hugging-face#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.