--- title: OpenJev emoji: 🎯 colorFrom: green colorTo: gray sdk: static app_file: index.html pinned: false --- Give OpenJev text or a screenshot and your options. Get back a decision and a score. A decision model has to be small enough and fast enough to sit inside a product, and still be general: any situation, any set of choices, given at request time. Small task-specific models are fast, but they are not general. OpenJev is a 27B model that answers in one pass on a single GPU and comes close to Jev: 84.0% against 85.4% on our 10,000-question text benchmark. OpenJev Flash 9B is the smaller one: about 1.6x faster, 79.4% on the same 10,000 questions, and on par with Cloudflare's Clef-Flash and ahead of Kev-9B and Nimble 9B on JevBench. [Get OpenJev](https://huggingface.co/openjev/openjev) · [Build with Loop](https://www.loopai.com/) > **New: deploy OpenJev on Runpod Serverless.** Runpod published a step-by-step guide for running OpenJev on one H100, no GPU of your own needed: [How to deploy OpenJev on Runpod Serverless](https://www.runpod.io/blog/how-to-deploy-openjev-on-runpod-serverless). ## Choose your model - [OpenJev](https://huggingface.co/openjev/openjev): BF16 weights for text and screenshots. - [OpenJev FP8](https://huggingface.co/openjev/openjev-FP8): GPU checkpoint for text and screenshots. - [OpenJev MLX](https://huggingface.co/openjev/openjev-MLX): 8-bit text decisions on Apple silicon. - [OpenJev MLX 4-bit](https://huggingface.co/openjev/openjev-MLX-4bit): text decisions on Apple silicon in about 15 GB. - [OpenJev GGUF](https://huggingface.co/openjev/openjev-GGUF): llama.cpp builds; Q4_K_M runs text decisions on a 24 GB card (RTX 3090 / 4090), Q8_0 for 40 GB and up. - [OpenJev Flash 9B](https://huggingface.co/openjev/OpenJev-Flash-9B): the 9B model, BF16 weights for text decisions, about 1.6x faster than OpenJev; on par with Clef-Flash, Kev-9B and Nimble on JevBench's public items. - [OpenJev Flash 9B MLX](https://huggingface.co/openjev/OpenJev-Flash-9B-MLX): the 9B model in 8-bit for Apple silicon. - [OpenJev Flash 9B MLX 4-bit](https://huggingface.co/openjev/OpenJev-Flash-9B-MLX-4bit), [FP8](https://huggingface.co/openjev/OpenJev-Flash-9B-FP8) and [GGUF](https://huggingface.co/openjev/OpenJev-Flash-9B-GGUF): the 9B for smaller Macs, vLLM without load-time quantization, and llama.cpp from 5.6 GB. **One server. Any model.** [openjev-server](https://huggingface.co/spaces/openjev/openjev-server) runs the same decision API on your own hardware, with OpenJev or any open model you choose, on a GPU or a Mac. `pip install` and one command to serve. Apache 2.0. The main model scores 84.0% on 10,000 text questions and 88.0% on 2,000 desktop screenshot next-action steps. [See the results and methods](https://huggingface.co/openjev/openjev#results). One forward pass per question with up to 52 choices. Open weights for research and non-commercial use under CC BY-NC 4.0. Commercial licence: [support@loopai.com](mailto:support@loopai.com). ## For enterprises **Policies change. The model does not need to.** The rules go in the request; the labels come back with probabilities, one pass each. Measured on generated cases with exact ground truth and one public dataset. [See the demos](https://github.com/loopai-hq/openjev-server/blob/master/demos/README.md). | **93.5%** | **200 of 200** | **95.0%** | |---|---|---| | policy decisions right, 200 cases under a nine-rule returns policy; 88.0% when it counts calendar days itself; majority baseline 42% | agent tool calls gated correctly under a seven-rule tool policy; 0 of 84 forbidden calls executed | prompt injections caught on 662 public messages; 97.7% recall at a 2% false-alarm rate | - **Runs where your data is.** Your GPU, your VPC or a Mac. Nothing leaves your network. - **Probabilities you can act on.** Calibrated scores: auto-approve above a threshold, send the rest to a person. - **One API, any model.** [openjev-server](https://huggingface.co/spaces/openjev/openjev-server) serves OpenJev or any open model behind the same interface. - **Commercial use.** The weights are CC BY-NC 4.0. For a commercial licence, email [support@loopai.com](mailto:support@loopai.com). - **Text and screens.** Text, JSON, DOM or a screenshot, one endpoint. 88.0% on 2,000 desktop next-action steps. - **Many questions, one call.** Route, score urgency and flag review in the same request; questions share the input but never see each other's answer. - **Fast enough for the request path.** About 125 ms per decision on one H100, about 150 ms on a Mac, up to 52 options in one pass. - **No lock-in.** Open weights you can audit, and the same API in front of any open model, so the vendor is never the bottleneck. **Why trust it.** Most small "decision models" only score well on their makers' own tests. On the same 10,000 public questions, same options, same order for every model: hosted Jev 85.4%, OpenJev 84.0%, the fine-tuned 9B we tested 75.7%. On par with the hosted model, eight points ahead of the small ones, and the method is public, so you can rerun it. OpenJev is an independent project, not affiliated with TypeSafe; Jev is their product.