Rizzo Flow
Open, local Jev-style server on llama.cpp that returns typed decisions from a fine-tuned Spark-X2.5-4B without generating tokens.
Overview
Rizzo Flow is a local-first implementation of the Jev programming model from Rizzo AI Academy: unstructured state in, typed probabilistic decisions out (yes/no, a choice among options, a rubric score, a number), with no generated text. It runs on llama.cpp, so it works on Apple, NVIDIA, AMD or Intel GPUs or CPU only, and serves both its own /v1/decisions schema and Jev's /v1/systemone, so code written for TypeSafe can point at localhost by changing a URL. Its model alias says explicitly that answers do not come from Jev.
The default model is Spark-X2.5-4B with the project's own LoRA fine-tune on public data. On the typed-decisions benchmark the README reports accuracy rising from 0.574 to 0.648 and ECE from 0.349 to 0.112 after fine-tuning, against Jev 1.13's 0.727 from its dataset card, and every published result names its dataset by hash. A Snake demo runs at about 150 ms per decision. The README warns probabilities are uncalibrated unless you calibrate on your own data. Italian documentation is also provided.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this convey urgency? | — |
| Choice | Which direction should the snake move? | — |
How it works
- Start the local server with a chosen model size and quantization
- POST state and typed questions to /v1/decisions or /v1/systemone
- Read typed answers with probabilities
- Calibrate thresholds on your own data before automating
Strengths
- Runs on almost any hardware via llama.cpp
- Drop-in URL swap for TypeSafe API code
- Open weights and fine-tune recipe
- Results tied to dataset hashes
Caveats
- Below Jev in accuracy per its own table
- Uncalibrated unless you calibrate
- Small 1.7B variant is much less accurate
At a glance
- Type
- Model
- Status
- beta
- Language
- Python
- Runtime
- HTTP server on llama.cpp
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
The open, local take on Jev: typed decisions from an LLM, without generating a…
wh000wh000/awesome-jev-liveLocal typed decisions from Spark-X2.5-4B over llama.cpp, serving both its own schema and Jev's `/v1/systemone`. Its `/v1/models` alias says in its description that it is not answered by Jev, and every published result names its dataset by hash.
MrJev/awesome-jev
Frequently asked questions
Is Rizzo Flow affiliated with TypeSafe?
No; it reproduces the interface pattern with an open model and does not use Jev.
Can my Jev client code use it?
Yes, it serves Jev's /v1/systemone so you can change the base URL to localhost.
How accurate is it?
The README reports 0.648 accuracy on typed-decisions versus 0.727 for Jev 1.13.
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