NeoHorse-Jev
TokenRhythm's NeoHorse repo: open 4B/9B agent models plus NeoHorse-Jev-4B, a prefill-only Choice/Noul/Score decision model.
Overview
NeoHorse is TokenRhythm's family of open-weight models for agent workflows. NeoHorse-1 provides 4B and 9B language models for tool use, coding and instruction following, released under Apache 2.0 with GGUF, quantized and MLX builds and a technical report on arXiv. NeoHorse-Jev-4B, released in September 2026, builds on NeoHorse-1-4B to make structured decisions with prefill-only inference.
Applications define questions and possible answers and use Choice to pick an action, Noul to check a condition and Score to rate. The README reports 77.70 across six text benchmark groups, the highest among four open-weight decision models with complete results in its comparison, and 83.26% mean accuracy on Nimble, VitaminC and MASSIVE, 11.50 points above the base model. It can be deployed with vLLM, SGLang or a native Python, CLI and HTTP runtime. It suits teams wanting a self-hosted decision model.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which action or tool should be selected? | — |
| Yes / no | Does this condition hold? | yes · no |
| Score | Rate the outcome on defined levels | — |
How it works
- Deploy NeoHorse-Jev-4B with vLLM, SGLang or the native runtime.
- Send state with application-defined questions.
- Get distributions, yes probabilities and expected ratings.
- Apply thresholds to route, select tools or control workflows.
Strengths
- Apache 2.0 open weights.
- Multiple deployment options including a native HTTP runtime.
- Published benchmark comparison and protocols.
Caveats
- Benchmarks are the authors' own evaluations.
- 4B model needs GPU resources for low latency.
Getting started
See jev/README.md in the repository for deployment with vLLM, SGLang or the native runtime.At a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- vLLM, SGLang, CLI, HTTP server
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Open alternative: Apache-2.0 4B decision model from TokenRhythm that answers `Choice`, `Noul` and `Score` questions via prefill-only inference on NeoHorse-1-4B, deployable with vLLM, SGLang or a native Python/CLI/HTTP runtime, scoring 77.70 across six text benchmark groups (highest among open-weight entries with complete results in its published comparison).
yibie/awesome-jev
Frequently asked questions
What is NeoHorse-Jev-4B?
A 4B decision model built on NeoHorse-1-4B that answers Choice, Noul and Score questions without text generation.
How well does it score?
The authors report 77.70 on a six-group text aggregate, highest among the open models they compared.
How can I deploy it?
With vLLM, SGLang, or the native Python, CLI and HTTP runtime.
Related projects
Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Three open 4B decision models for English, Turkish and German that serve the Jev /v1/systemone API on a CPU in about 3 GB of RAM.
ServicebetaPython