Open-source Jev alternatives: System One models you can self-host

A wave of open-source alternatives to TypeSafe Jev, the hosted System One model that returns probabilistic answers to typed questions rather than generated text, has emerged within days of its September 15, 2026 launch. The models — including Laya, Kev, Decider, Von and OpenThai-SystemOne, most under Apache 2.0 — can be self-hosted and several serve the same POST /v1/systemone endpoint, letting existing Jev client code switch with only a base URL change. Upstream Laya reports 0.425 on Banking77's 77 intents versus Jev's 0.870, though open models close much of the gap after fine-tuning on local data.

TypeSafe Jev is the hosted model that introduced System One models on September 15, 2026: you send a state and typed questions, and it returns answers with probabilities instead of generated text. It is in early access and billed per token third-party write-ups cite $0.042 per million input tokens . Within days, open-source Jev alternatives appeared that do the same job on your own hardware. This post lists the ones worth testing, which of them accept Jev API requests unchanged, and what it takes to self-host one. If you already call Jev, start with a model that serves the same POST /v1/systemone endpoint, so existing client code only needs a new base URL: | Model | Size | License | Serves the Jev API | Good first choice when | |---|---|---|---|---| | Laya https://laya-ai.com/models | 322M / 421M | Apache 2.0 | Yes, through laya-serve | You want a small model that runs on CPU, or 100+ languages | | Kev https://laya-ai.com/system-one-models/kev | 0.8B to 27B | Apache 2.0 | Yes, the TypeSafe SDK works unchanged | You have a GPU and want the closest drop-in | | Decider https://github.com/Mapika/decider | 2B / 4B / 35B MoE | Apache 2.0 | Yes | You serve with vLLM and want many-option choices | | Von https://laya-ai.com/system-one-models/von | 395M | Apache 2.0 | Yes | You need low latency on modest hardware | | OpenThai-SystemOne https://github.com/iapp-technology/openthai-systemone | 0.8B | Apache 2.0 | Yes | Your input is Thai | Other open options use their own APIs: Bespoke Nimble 9B , GLiNER2.5-Decide 340M , Together AI's Tev1-4B https://laya-ai.com/system-one-models/tev1 , and the training-free AnyJev and SemIf. See the full System One model comparison https://laya-ai.com/system-one-models for all twelve. These projects put a Jev-style HTTP API in front of an open model, so you can switch models later without touching client code: | Server | Language | Notes | |---|---|---| | laya-serve | Python | Ships with the upstream Laya package. See Self-host Laya https://laya-ai.com/guides/self-host-laya . | | Arbiter https://github.com/0xBakeer/arbiter | Python | Routing, batching, metrics and a built-in playground | | sys1 https://github.com/alvarobartt/sys1 | Rust | candle-based, token batching, CPU, CUDA and Metal | | ollaya https://github.com/ollaya-dev/ollaya | Rust | Ollama-style CLI and daemon; pulls models by name | | laya-server https://github.com/1Panel-dev/laya-server | TypeScript | Docker image with a web console and API keys | For the two Rust servers side by side, see Run Laya in Rust https://laya-ai.com/runtimes/rust . TYPESAFE BASE URL=http://localhost:11435 . With plain HTTP, the request body stays the same: curl -X POST http://localhost:8080/v1/systemone \ -H 'Content-Type: application/json' \ -d '{"state":{"message":"I was charged twice"},"questions":{"refund":{"type":"noul","instructions":"Does the customer ask for a refund?"}}}' Hosted Jev still leads out of the box on several published comparisons, especially on tasks with many labels. On Banking77, with 77 intents, upstream Laya reports 0.425 at default settings against Jev's 0.870. Open models close much of the gap after fine-tuning on your own data, and they win on cost, privacy and latency when you run them close to your application. See Laya vs Jev https://laya-ai.com/laya-vs-jev for the details and caveats. Benchmarks in this space are young and mostly self-reported on different datasets, so treat any single number as a starting point, not a verdict. No. Jev's weights are not public. The Jev alternatives above are independent open models, mostly Apache 2.0, that answer the same kind of typed questions. Several of them also accept the same POST /v1/systemone requests. The encoder models are the lightest: Laya 322M and 421M parameters , Von 395M and GLiNER2.5-Decide 340M . See Laya on CPU https://laya-ai.com/guides/laya-cpu-performance for measured latency. Yes, with a server that speaks the Jev API. Kev and ollaya say the SDK works unchanged, and laya-serve says an existing Jev client only needs a new base URL. Last verified: September 25, 2026. I maintain laya-ai.com https://laya-ai.com , an independent Laya resource site; this article first appeared there as Jev alternatives https://laya-ai.com/jev-alternatives . The site also keeps an auto-updated list of System One models on GitHub https://github.com/yanng981/awesome-system-one . It was drafted with AI assistance, and every fact was checked against the project pages linked above.