https://github.com/feder-cr/jev
jevos: an open-source alternative to Jev and Microsoft-Decision-1. Yes/no decisions, choices and scores with probabilities, on your laptop's CPU.
https://github.com/feder-cr/jev
ai classification cpu-inference decision-model edge-ai gguf jev jev-alternative llama-cpp llm llm-inference local-ai local-llm microsoft-decision microsoft-decision-1 offline on-device-ai open-source-alternative self-hosted yes-no
Last synced: 29 days ago
JSON representation
jevos: an open-source alternative to Jev and Microsoft-Decision-1. Yes/no decisions, choices and scores with probabilities, on your laptop's CPU.
- Host: GitHub
- URL: https://github.com/feder-cr/jev
- Owner: feder-cr
- License: mit
- Created: 2024-08-15T14:24:48.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2026-10-16T23:53:22.000Z (about 1 month ago)
- Last Synced: 2026-10-17T12:25:13.924Z (about 1 month ago)
- Topics: ai, classification, cpu-inference, decision-model, edge-ai, gguf, jev, jev-alternative, llama-cpp, llm, llm-inference, local-ai, local-llm, microsoft-decision, microsoft-decision-1, offline, on-device-ai, open-source-alternative, self-hosted, yes-no
- Language: C++
- Homepage: https://huggingface.co/feder-cr/jev
- Size: 17.1 MB
- Stars: 1,346
- Watchers: 25
- Forks: 163
- Open Issues: 5
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
- awesome-jev-tools - jev - SDK & Decision Frameworks: jevos is an open-source, Jev-compatible alternative to TypeSafe's Jev for yes/no decisions that runs entirely on a laptop CPU. Send text plus a yes/no question over HTTP and get back P(yes) from a single forward pass of a 1B model โ no text generation, no GPU required. (Full list / Infra / SDKs / Integrations)
- awesome-jev-projects - jev - jevos is an open-source, Jev-compatible alternative to TypeSafe's Jev for yes/no decisions that runs entirely on a laptop CPU. Send text plus a yes/no question over HTTP and get back P(yes) from a single forward pass of a 1B model โ no text generation, no GPU required. (SDK & Decision Frameworks)
- awesome-jev - **jevos** - Open-source, CPU-only alternative to Jev for yes/no decisions: MiniCPM5-1B cut to 17 layers with a one-logit output head, quantized to a 619 MB GGUF, one forward pass on llama.cpp, no GPU. ยท `C++` (Open-source alternatives)
- awesome-jev - jevos - MiniCPM5-1B cut to 17 layers with a one-logit output head, quantized to a 619 MB q4_k_m GGUF for CPU-only inference under llama.cpp behind Jev's `/v1/systemone` wire format; answers noul (yes/no) questions only, choice and score return 422; reports 0.815 accuracy on 2,000 held-out yes/no rule questions against 0.927 for hosted Jev, and 54-220 ms per request on a 16-thread laptop CPU versus Jev's hosted 344-345 ms. (Open reproductions and research)
- awesome-jev-gallery - jevos - 1B cut to 17 layers and served as a 619 MB GGUF on CPU-only llama.cpp; answers Noul (yes/no) questions over Jev's `/v1/systemone` wire in 54 to 220 ms, 0.815 accuracy versus Jev's 0.927. [](https://github.com/feder-cr/jev) (๐งช Open Source)
- awesome-jev - jevos - Self-hosted, Jev-compatible `/v1/systemone` server answering yes/no (Noul) questions from a CPU-only 1B GGUF model (MIT); independent of hosted Jev. [Project guide](community/projects/tools/jevos.md). (Community projects / Developer projects and integrations)
- awesome-jev - jev - Jevos is an open-source alternative to Jev for yes/no decisions that runs on your laptop. (Research, Evals, and Reimplementations)
- awesome-decision-models - jevos - 1B decision model for CPU-only laptops: MiniCPM5 cut to 17 layers with a one-logit head, GGUF q4_k_m at 619 MB, running on llama.cpp with no GPU, about 54 ms per short request. Speaks Jev's `/v1/systemone` wire format for Noul (yes/no) questions only; Choice and Score return 422. (Open Models / Community models)
- awesome-jev - jevos - cr/jev?style=flat-square&label=%E2%98%85) - Open alternative: MIT-licensed 1B model (MiniCPM5 cut to 17 layers with a one-logit head, GGUF q4_k_m at 619 MB) that answers only `Noul` yes/no questions on Jev's own `/v1/systemone` wire format, running CPU-only via llama.cpp at 54 ms short / 220 ms long on a laptop Core Ultra 7 255H against Jev's hosted 344/345 ms; `Choice` and `Score` return 422, and it scores 0.815 accuracy on 2,000 unseen policy yes/no questions against Jev's 0.927. (Full list / Calibration & Research)
- awesome-jev - jevos - Serves `/v1/systemone` and `/v1/models` for yes/no questions only โ other types are refused โ from a GGUF model through llama.cpp on CPU, offline once the file is downloaded. (Open Models & Reproductions)
- awesome-typesafe-jev - jevos - licensed 1B yes/no model released as a q4_k_m GGUF that runs on CPU with llama.cpp behind a Jev-compatible `/v1/systemone` endpoint; it answers only Noul questions, and Choice and Score requests return 422. The authors report 0.815 accuracy against 0.927 for hosted Jev on 2,000 questions from unseen policies, but the training code and evaluation set are not published, so that result cannot be reproduced from the repository. (Community projects / Evaluations and independent research)
- awesome-jev - jevos - 09-28 | Local 1B CPU decision model with Jev-compatible Noul responses; it returns P(yes) and explicitly rejects Choice and Score. | (Added today / Jev-like models (80))
README

**Decisions on a laptop CPU in 28โ130 ms.** Send a text and a question (yes/no, multiple choice or a
score) and get back the probability of each answer, in TypeSafe Jev's API, on your own machine.
## Quickstart
From the [release](https://github.com/feder-cr/jev/releases/tag/jevos-v5), download the archive for your
system and `jevos-v5-openvino-int8.zip`, then:
```bash
tar -xzf jev-linux-x64.tar.gz # Windows: unzip jev-windows-x64.zip
cd jev
unzip ../jevos-v5-openvino-int8.zip # creates model/
./jev serve # Windows: jev.exe serve
```
The model is also on Hugging Face, [feder-cr/jev](https://huggingface.co/feder-cr/jev):
`hf download feder-cr/jev --include "model/*" --local-dir jev` puts it in `jev/model/`.
```bash
curl http://127.0.0.1:8017/v1/systemone -H 'Content-Type: application/json' -d '{
"model": "jev-latest",
"state": "My invoice shows a charge I do not recognize.",
"questions": {"billing": {"type": "noul", "instructions": "Is this a billing problem?"}}}'
```
```json
{"model": "jevos-v5", "answers": {"billing": {"type": "noul", "noul": 0.97}}, "usage": {"input_tokens": 28, "output_tokens": 0}}
```
One binary, CPU only, no Python. The release and the Hugging Face repo also have the model as GGUF for
llama.cpp.
## Benchmarks

Same questions for every system, through the same HTTP client. Latency is the median of 10 requests (for
jevos, of 40 over four sessions) on an Intel Core Ultra 7 255H laptop, 16 threads, each text read from
scratch. No task text was used to train jevos; all seven sets were among those used to choose the released
checkpoint.
| | **jevos-v5** | Jev | Laya |
|---|:---:|:---:|:---:|
| Yes/no questions | โ | โ | โ |
| Multiple choice | โ | โ | โ |
| Scores | โ (early) | โ | โ |
| Runs on | your machine | cloud | your machine |
| Cost | free | per token | free |
| Context | 8,192 tokens | not stated | 512 tokens |
## API
`POST /v1/systemone`, in TypeSafe Jev's wire format: code written for Jev's SDK works unchanged for yes/no
questions. One request can ask several questions about the same text, which is read once:
```json
{
"model": "jev-latest",
"state": {"item": "wireless mouse", "customer_message": "The box arrived empty. This is the second time!"},
"questions": {
"refund": {"type": "noul", "instructions": "Our policy refunds items reported missing within 30 days of delivery. Should this customer get a refund?"},
"team": {"type": "choice", "instructions": "Which team should handle this message?",
"criteria": {"billing": "payments, refunds", "shipping": "deliveries, missing parcels", "tech": "a product that does not work"}},
"anger": {"type": "score", "instructions": "How angry is the customer?", "criteria": ["calm", "annoyed", "angry", "furious"]}
}
}
```
A `noul` returns P(yes); a `choice` the most probable option; a `score` the expected level. Choices and scores
also return a probability per answer and a `confidence`: when it is low, send the case to a person. Put the
rule in the question, and do sums in code.
`GET /health` reports readiness and the SHA-256 of the model files; `./jev decide request.json` answers a
request file without a server.
## Options
| Option | Default | |
|---|---|---|
| `--threads` | all logical CPUs | fewer if other heavy apps are running |
| `--host`, `--port` | `127.0.0.1`, `8017` | where the server listens |
| `--model-dir` | `model` beside the binary | the model folder |
With `JEV_API_KEY` set, every call but `/health` needs `Authorization: Bearer `. About 1 GB of memory,
1.4 GB with the text cache full. Every option is described at the top of [src/main.cpp](src/main.cpp).
## Build from source
```bash
python -m pip install -r requirements.txt # OpenVINO's SDK, CMake, Ninja, the tests' packages
python scripts/build.py # dist/jev; on Windows, from a Visual Studio developer prompt
python tests/check.py # with the model in dist/jev/model
```
## More
Guides and measurements are in the [wiki](https://github.com/feder-cr/jev/wiki). Built together with
[Loris Salsi (@LosaLosSantos)](https://github.com/LosaLosSantos).
