Nimble is the best of the open source jev alternatives right now, with OpenJev the strongest pick for teams that want a self-hosted drop-in API and Jeff the lightweight option for pure classification — and with TypeSafe pausing new Jev signups on 22 September 2026, the open ecosystem has gone from curiosity to necessity in a week. Jev, TypeSafe AI's "System One" decision model, returns typed answers — a choice, a score, a yes/no — with calibrated probabilities instead of prose, and the open-source community has been busy rebuilding that idea on open weights. This ranking covers the three projects worth your time, what each actually is, and which fits which job.
📺 Watch: Run Jev AI For FREE, Here’s How!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
Context for the ranking: Jev launched on 15 September 2026, and The Register's 23 September 2026 write-up describes its three primitives — Choice, Score and Noul (a calibrated yes/no probability) — priced at $0.042 per million input tokens with output free. Direct signup at console.typesafe.ai opened to everyone on 20 September with $5 of credit, then TypeSafe paused new signups on 22 September to protect service for existing users. That pause is why the projects below matter: they are how you get typed decisions today without waiting on a signup queue. If your goal is free access to Jev itself rather than a replacement, the how to use Jev AI for free guide covers those routes separately.
The Best Open Source Jev Alternatives, Ranked
1. Nimble (Bespoke Labs) — the open recipe closest to Jev's accuracy
Nimble is Bespoke Labs' open framework and model for typed decisions from text, and it currently sits closest to Jev on published agreement numbers. Per the project's GitHub README, Bespoke-Nimble-9B is a LoRA fine-tune (rank 16) of Qwen3.5-9B that works by logit scoring: it processes the context and schema once, then reads the scores of the allowed answer tokens directly — no generated JSON, no reasoning text, one step per decision.
The numbers, from the same README: on 324 held-out examples, Bespoke-Nimble-9B agrees with reference labels 90.12% of the time, against 93.21% for Jev 1.13.0 and 66.36% for untuned Qwen3.5-9B — and it beats the much larger Qwen3.8-27B (84.88%) on the same set. The repo saw its latest checkpoint update on 24 September 2026, so it is being actively developed. Requirements are honest rather than trivial: Python 3.12, Apple Silicon with MLX or an NVIDIA GPU with BF16 support, roughly 18GB for unquantised weights, and an 8,192-token context limit. If you can meet that hardware bar, Nimble is the pick — a genuinely open recipe (training approach included) whose accuracy lands within about three points of the hosted product it replaces.
If you want to turn decision models like these into working automations — routing, lead scoring, agent pipelines that run without you — the AI Profit Boardroom has the builds and weekly live coaching calls to get you shipping → Build your first decision pipeline inside. Prefer 1-on-1 help on the SEO side? Book a free SEO strategy session.
2. OpenJev — the self-hosted decision API for production stacks
OpenJev takes the other route: instead of one fine-tuned model, it turns open-weight LLMs you already run into a typed probabilistic decision service, no fine-tuning required. It reimplements the familiar primitives — Choice for categorical selection, Noul for binary judgments, Score for ordinal ratings — and adds first-class abstention options (UNKNOWN, OUT_OF_SCOPE, HUMAN_REVIEW), which production triage systems genuinely need. Under the hood it evaluates candidate log-likelihoods on a vLLM or SGLang endpoint with constrained decoding, then applies temperature-scaling calibration to keep the probabilities honest; the repo claims 40–120ms end-to-end on local vLLM deployments.
Model support spans tiers from Qwen3-1.7B/4B and Gemma 4 for speed up to DeepSeek-V4.1-Flash for complex calls. The catch is the licence: OpenJev is source-available under PolyForm Noncommercial 1.0.0, free for personal and academic use, but commercial deployment requires a written licence from the maintainers. That makes it the strongest architecture on this list and the most constrained legally — read the licence before you build revenue on it. The dedicated OpenJev guide on this site goes deeper on setup if it fits your case.
3. Jeff — the tiny classifier that speaks TypeSafe's API
Jeff is the minimalist of the three: per DataCamp's round-up of open-source Jev alternatives, it packages a 400M-parameter GLiFormer classifier behind TypeSafe's System One API shape, serving requests at a local /v1/systemone endpoint so existing TypeSafe SDK clients can connect by changing the base URL. That makes it the cheapest migration path in engineering terms — point your client at localhost and keep your code.
The trade-off is scope. A 400M classifier is built for classification-shaped jobs — ticket routing, labelling, triage, policy checks — not for the harder judgment calls a 9B fine-tune or a large open model can make. Used inside that lane, small is a feature: minimal hardware, fast responses, no licence drama. Used outside it, you will feel the ceiling quickly.
Open Source Jev Alternatives Compared
| Project | What it is | Published accuracy claim | Licence | Best for |
|---|---|---|---|---|
| Nimble | Qwen3.5-9B LoRA fine-tune + open recipe, logit scoring | 90.12% vs Jev 1.13.0's 93.21% (324 held-out examples) | Open recipe and weights on GitHub | Closest open accuracy to Jev, local-first teams |
| OpenJev | Decision API layer over open models via vLLM/SGLang | 40–120ms local latency; calibrated via temperature scaling | PolyForm Noncommercial 1.0.0 (commercial needs a written licence) | Self-hosted production-style stacks, non-commercial use |
| Jeff | 400M GLiFormer classifier behind TypeSafe's API shape | Classification-scope only | Open source | Ticket routing, labelling, triage on small hardware |
One honest caveat on comparing accuracy claims across these open source Jev alternatives: each project publishes its own evaluations on its own test sets, so treat the numbers as each maintainer's claim rather than one benchmark. The pattern is still clear — the gap between open and hosted has narrowed to a few points on the published sets.
How to Choose Between Them
Choose by constraint, not by vibes. If accuracy on general typed decisions is the priority and you have Apple Silicon or an NVIDIA card, run Nimble. If you need an API your whole stack can call, you run open models already, and your use is personal or academic, OpenJev is the most complete architecture. If your jobs are classic classification and your hardware is modest, Jeff gets you moving in an afternoon. And if what you actually want is Jev itself at zero cost, check the current free routes covered earlier and the paid maths in Jev pricing before committing to self-hosting — at $0.042 per million input tokens, the hosted product is cheap enough that open source has to win on access, control or privacy rather than price.
For inspiration on what to build once you have any of these running, the Awesome Jev GitHub list catalogues 100+ decision-shaped use cases with public repos behind them, and the Jev automation guide shows how typed decisions slot into larger agent workflows. Decision models are routers, not writers — their value shows up when something bigger, like the Agent OS, is doing the orchestration around them. And if you are wondering which large model should sit behind OpenJev for the hard calls, the Goldie Bench write-up covers how the current brains compare in hands-on tests.
📺 Watch: How to Setup Jev AI + Claude Code
Open Source Jev Alternatives FAQs
Is there a fully free and open alternative to Jev?
Nimble is the closest: open recipe, open checkpoint, runs locally with no TypeSafe key. OpenJev's code is source-available but non-commercial by default, and Jeff is open but classification-scoped. None requires a TypeSafe account.
How close are open source Jev alternatives to the real thing?
On Nimble's published 324-example evaluation, the gap is about three points of label agreement (90.12% vs 93.21% for Jev 1.13.0). For classification-style workloads the practical gap can be smaller; for subtle judgment calls the hosted model still leads.
Why would anyone self-host when Jev costs $0.042 per million tokens?
Access and control. TypeSafe paused new signups on 22 September 2026, hosted services can change terms, and some workloads cannot leave your infrastructure. Self-hosting trades convenience for a service nobody can switch off under you.
Can these alternatives handle Jev's Noul primitive?
OpenJev implements Noul directly alongside Choice and Score. Nimble covers the equivalent ground with enum and true/false questions returning probabilities. Jeff handles binary decisions within its classification scope.
Verdict: The Open Bench Is Real Now
A week ago, open source Jev alternatives were a hedge; after the signup pause they are the way many teams will actually experience typed decision models. Nimble proves the accuracy gap is closable with a 9B model and a good recipe, OpenJev proves the API layer can be rebuilt on open weights, and Jeff proves the small end of the market needs almost nothing. Pick the one that matches your constraint, start with a low-stakes routing job, and let the results argue for expansion.
If you want the shortcut — working decision-model automations, agent templates and a community shipping this stuff daily — check out the AI Profit Boardroom → Join the builders inside. And for a personal plan on ranking with AI-driven SEO, book a free SEO strategy session.
