Rizzo Flow
Local llama.cpp Jev-shaped server: Spark-X2.5 LoRA, POST /v1/systemone on :8017, zero answer tokens, independent of TypeSafe.
Overview
Rizzo Flow (github.com/Rizzo-AI-Academy/rizzo-flow, Apache-2.0) is an independent local server that maps unstructured state plus Choice, Score, and Noul questions to probabilities with zero generated answer tokens on llama.cpp. README states it is not affiliated with TypeSafe and does not reproduce proprietary Jev architecture; it follows the interface pattern with Spark-X2.5 plus a September 2026 LoRA fine-tune for typed decisions. uv sync, rizzo download, and rizzo serve default to http://127.0.0.1:8017 with a playground UI. curl examples hit POST /v1/systemone; rizzo decide runs one-off JSON. Backends include Metal, CUDA, Vulkan, ROCm, SYCL, and CPU. Fine-tune tables on LocalLLaMA/typed-decisions show accuracy 0.648 vs 0.574 base Spark at Q8_0, with explicit uncalibrated probability warnings unless you calibrate locally.
Problem: Developers want Jev-shaped probabilities on localhost without api.typesafe.ai bills, but chat-json bridges and unrelated classifiers do not ship open decision weights.
Built for: Builders with Python 3.11+, uv, and a GPU or CPU backend who want POST /v1/systemone and /v1/decisions against Spark-X2.5 on llama.cpp.
First indexed on Jev Directory: 2026-09-25
Creator and team
- Name
- Rizzo AI Academy
- Handle
- @Rizzo-AI-Academy
How Jev is used
- Role in the product flow
- Local llama.cpp inference exposing TypeSafe-compatible System One HTTP on port 8017
- Primitives
- ChoiceScoreNoul
- State in
- JSON state string plus questions map with type choice, score, or noul, instructions, and criteria per README curl and playground examples.
- Decision out
- Parallel typed answers with probability vectors; README emphasizes 0 generated tokens on the decision path.
- git clone, uv sync --locked, uv run rizzo download for runtime plus Q8_0 weights
- uv run rizzo serve for daemon and playground on :8017
- POST /v1/systemone or /v1/decisions from TypeSafe-shaped clients
- Optional rizzo decide path for single JSON files without keeping the server up
Rizzo Flow is the open-weights counterpoint to hosted System One: same HTTP verbs, different training story than Jared Palmer Kev or Together tev1. githubnext-localjev still prompts a chat model for JSON; Rizzo runs a dedicated fine-tuned Spark head on llama.cpp. wfzyx-von and jaredpalmer-kev are alternative checkpoints; theoleecj-semif documents another local server narrative. nokia-applied-research-anyjev fits heads on models you already serve on vLLM. Treat README calibration disclaimers seriously before you wire refund gates.
Sourced performance claims
- typed-decisions benchmark table documents Rizzo Flow 4B fine-tune at accuracy 0.648, KL 0.452, Brier 0.205, ECE 0.112 vs base Spark 0.574 accuracy at same Q8_0 GPU setup.Source: github.com/Rizzo-AI-Academy/rizzo-flow README fine-tuning section
- Badge and Snake demo cite about 50 ms per decision at Q8_0 on RTX 5060 Ti and about 150 ms round trip in recorded Snake gameplay with caveats on hardware.Source: github.com/Rizzo-AI-Academy/rizzo-flow README
- Public GitHub repo had about 489 stars when this listing was drafted.Source: GitHub API September 2026
Features and stack
Features
- POST /v1/systemone and /v1/decisions on llama.cpp with playground UI
- Spark-X2.5 4B and 1.7B downloads with resumable rizzo download
- September 2026 LoRA fine-tune with published typed-decisions metrics
- Snake and ticket demos at real-time latencies with documented caveats
- Explicit not affiliated with TypeSafe disclaimer and uncalibrated probability notes
Stack
- Python 3.11+
- uv
- llama.cpp
- Spark-X2.5
- LoRA fine-tune
Pricing: Apache-2.0 open source; inference cost is your local hardware and electricity.
Links
FAQ
- Is Rizzo Flow official TypeSafe Jev?
- No. README labels the project independent, inspired by SemIf, with open Spark weights and interface compatibility only.
- Which endpoint should TypeSafe SDK clients use?
- Point TYPESAFE_BASE_URL at http://127.0.0.1:8017 and call the same System One shapes documented in README curl examples.
- Are the probabilities production calibrated?
- README states they are uncalibrated unless you calibrate on your data. Compare ECE on your holdout before hard thresholds.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev & TypeSafe
A primer for builders who need to know what Jev is, who it is for, and when TypeSafe docs should take over from this directory.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.