Jeff
0.8B open System One model with task adapters that makes millisecond decisions in front of a larger local model like Qwen3.8-27B.
Overview
Jeff is a small (0.8B) open System One model meant to sit in front of a strong local model such as Qwen3.8-27B: the big model writes and plans, and Jeff makes the quick decisions, deferring to the big model only when unsure. Version 1.3 has an adapter-first base, 15 task adapters (guard, triage, support intents, tool selection, grounding, voice navigation, spam, legal clauses, emotion and more), GGUF files for llama.cpp, and Jeff-Code, which places Jeff inside a coding agent.
The README reports, across eight adapters on the same test rows, 94.6% mean accuracy for Jeff alone versus 86.6% for the 27B, at 0.25 s versus 8.1 s per decision on an Apple M4 Max, and 95.0% with the 27B as fallback. Jeff-Code is reported as 32% less time per task at a similar pass rate over 1,242 tasks. The yibie list also records about 22 ms per decision, a panel measuring Jev at 0.828 accuracy and 0.053 ECE, and an explicit note that no statistical significance is claimed. Licences differ by adapter; some are CC BY-NC 4.0.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this a prompt injection or jailbreak? | yes · no |
| Choice | Which tool should an agent call? | — |
| Yes / no | Is the answer supported by the sources? | yes · no |
How it works
- Load the Jeff base with the needed adapters
- Send /v1/systemone-shaped requests
- Act on confident answers in milliseconds
- Pass unsure questions to a larger local model
Strengths
- Reported faster and more accurate than a 27B on most adapter tasks
- Small memory footprint (about 2 GB with all adapters)
- Runs locally via llama.cpp, MLX or GPU
- Same request format as Jev
Caveats
- Adapter licences differ; some are non-commercial
- v1.3 adapters only work on the v1.3 base
- Author states no statistical significance is claimed
Getting started
Download the base and adapters from Hugging Face (or GGUF files for llama.cpp) and follow the Quick start at jeffhub.ai.At a glance
- Type
- Model
- Status
- beta
- Language
- Python
- Runtime
- local server, llama.cpp, MLX
- Has code
- yes
Use cases and integrations
How curated lists describe it
Open alternative: MIT-licensed Qwen3.5 and Gemma 4 fine-tunes answering `choice`, `noul` and `score` on the same `/v1/systemone` format at about 22 ms per decision, published with a panel that measures Jev itself at 0.828 accuracy and 0.053 ECE while stating it claims no statistical significance.
yibie/awesome-jevQwen3.5 and Gemma 4 fine-tunes for zero-shot classification that take Jev's request shape. The author reports about 22 ms per decision on an RTX PRO 6000 and 28 ms on an M4 Max under MLX.
MrJev/awesome-jev
Frequently asked questions
What is Jeff for?
Making fast typed decisions in front of a larger local model, which only handles questions Jeff is unsure about.
How fast is it?
The README reports 0.25 s mean per decision on an M4 Max versus 8.1 s for a 27B.
Can I use it commercially?
Depends on the adapter; some are CC BY-NC 4.0 because of their source data.
Related projects
Clojure System One engine on Jolt that routes typed questions to Laya encoders, a llama.cpp 'thinker', Clef or TinyDecide.
ModelresearchClojurevLLM Semantic Router's Apache-2.0 decision models (0.6B to 27B) answering up to 64 typed questions per input in one forward pass.
ModelbetaPythonMangaba 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.
ServicebetaTypeScript