Article
CommunityJev Guides & Articles
Jev Explained: RLCD Internals and Local Runners
accept99 explains how Jev differs from chat models and how to run Jev-like decision models for free: Ollama serves nimble 9B and tev1 4B and 0.8B over /v1/systemone, while the open-source Laya project offers pip-installable weights and an MLX build for Macs.
- Format
- Article
- Reading time
- 2 min
- Published by
- Community
- Author
- accept99
- Last verified
- Oct 2026
- Languages
- Shell
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- The page calls Jev a machine-native, large programmable model, not a decision model, per TypeSafe CEO Diogo Almeida.
- Ollama hosts nimble at 9B plus tev1 at 4B and 0.8B, all local and free, served over a Jev-compatible /v1/systemone endpoint.
- A Pac-Man test measured about 91 ms per decision call through Ollama on an Apple M5 Max.
- Laya installs with pip install laya; its serve extra starts FastAPI on POST /v1/systemone, and Laya-MLX runs natively on Macs under 1 GB of memory.
- The page warns that Tev1 logprobs are model preferences, not calibrated confidence, and cites SOTAAZ measurements where Nimble overestimated by about 21 points.
Watch out
The author asks readers to verify domains, model names, and gateway facts against TypeSafe and platform documentation. The 91 ms figure and the Laya-MLX latency are single anecdotal measurements, and the page stresses that Tev1 and Nimble probabilities stay uncalibrated until shadow-mode retesting on your own data.
Questions about this entry
Short answers drawn from the linked project or page.
It is the second article in a Jev series: why the model differs from chat systems, how to call it correctly, and two routes for running Jev-like decision models locally with Ollama or the open-source Laya project.
The page says RLCD, Reinforcement Learning for Calibrated Decisions, targets the program-in-the-loop use case of reliable structured outputs for software, as opposed to RLHF for chat products and RLVR for benchmark-verifiable reasoning and code.
Choice corresponds to a switch over an enum, Noul corresponds to an if condition, and Score corresponds to sorting and threshold comparisons. Noul is named after the Bernoulli distribution and returns a scalar probability between 0 and 1.
Ollama serves nimble at 9B plus tev1 at 4B and 0.8B; ollama pull nimble installs one, and requests go to /v1/systemone in a Jev-compatible format that returns per-option probabilities and confidence.
Yes. The open-source Laya model installs with pip install laya and a FastAPI server with pip install "laya[serve]", exposing the same /v1/systemone path; Laya-MLX runs on Macs without PyTorch using under 1 GB of memory.
A common community split: 0.8 and above auto-executes, 0.5 to 0.8 handles conservatively with truncation or summary, and below 0.5 deletes, falls back, or escalates to a stronger model or a person. The article stresses recalibrating these thresholds on your own data first.
The page says Tev1's README calls its logprobs model preferences rather than calibrated confidence, and third-party SOTAAZ tests showed Nimble overestimating by about 21 points on JudgeBench and WinoGrande, so the thresholds need shadow-mode recalibration.
More Jev resources like this
The Ollama build of iapp's OpenThai-SystemOne: a 0.8B Thai and English decision model whose v0.3 weights are fine-tuned for Ollama's /v1/systemone letter readout, reporting a 74.3 macro across 13 public subsets for Q8_0.
Jev Models & Reimplementations#community#open-models#local
Community
TurinTech's independent GGUF conversion of the 27B Kev decision model for llama.cpp: the default IQ4_XS file is 13.91 GB and keeps 97.8% of the BF16 original's answers over 2,000 test decisions, with zero changed confident answers.
Jev Models & Reimplementations#community#open-models#local
Community
An 8-bit MLX conversion by SirSahOl of chaoliangUNSW/Jev-Style-0.8B-Decision-v3 built for Apple Silicon native GPU inference: about 990 MB of active VRAM, a 262,144-token context and an average of 8.25 bits per weight.
Jev Models & Reimplementations#community#open-models#local
Community
A Q8_0 GGUF of knowlet's Gemma-4-12B-Unified-System-One with a native System One decision runtime and an independently fitted GGUF temperature, shipped with an F16 mmproj for the model's native image and audio projections.
Jev Models & Reimplementations#community#open-models#quantized
Community
A 27B decision model from StartLux Labs that takes text, JSON or images up to 262,144 tokens and returns a probability per option through the TypeSafe /v1/systemone format, reporting Decision Index 0.2.1 63.88 and 208 of 231 on JevBench public.
Jev Models & Reimplementations#community#open-models#local
Community
The Q8_0 8-bit GGUF build of StartLux-Decision-v1-27B, 28.60 GB for llama.cpp, served through a decision wrapper that the card measures at 100% same decisions as the original and 209 of 231 on JevBench public.
Jev Models & Reimplementations#community#open-models#quantized
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.