Jev Reality Check: TypeSafe's Model Stays Closed, but Open Jev-Class Models Now Run Locally

Last updated: September 2026

No, not TypeSafe's Jev: it is still a closed, hosted API with nothing to download, and Ollama has no native support for it. What you can run is an open Jev-class model. Several now answer the same typed questions on your own hardware, and some accept TypeSafe's request format, so the official SDK works against a server on your network.

Updated September 25, 2026. Three things moved since our launch-day pass on September 15. An open field appeared, and an independent benchmark, JevBench, scored 93 systems on September 24 with an Apache 2.0 model at the top of its official ranking. TypeSafe dropped its waitlist on September 20, then paused new signups two days later under demand. And the do-it-yourself method this page described on day one now has a score: a frozen Gemma 4 12B read that way, plus one calibration setting, finished 1.5 points behind Jev.

Can you run Jev itself locally? Still no

TypeSafe's Jev remains a closed-weight model served from the United States, with no download, no on-premises build and no published plan for either. Jev 1.13 is still the current version.

Access loosened, then tightened. TypeSafe dropped the waitlist on September 20 and paused new signups on September 22 because demand outran capacity; existing accounts keep working, and on September 25 its homepage offered only a sign-in link. Vercel's AI Gateway and Cloudflare Workers AI serve Jev without a TypeSafe account, and OpenRouter's listing prices it at the same $0.042 per million input tokens, with output free.

The privacy terms have not changed. TypeSafe's privacy policy, last updated November 19, 2025, says it will not train or fine-tune models on your input and will not disclose that input to anyone but its service providers. Every decision still leaves your network, though.

The "can't hallucinate" claim also means what it meant on launch day. Jev cannot return an answer outside the options you define, but it can still pick the wrong one, and TypeSafe's launch post says its zero-percent hallucination figure is not an empirical measurement.

The open Jev-class models you can run today

Dozens of open projects now answer Jev's three question types (a choice, an ordered score, a yes/no probability), and the best of them come within a few points of Jev on routine decisions. Most of the systems in JevBench's latest run are open rebuilds of this kind.

How JevBench measures them

JevBench is Benchmark Heaven's own benchmark and is not affiliated with TypeSafe. Its September 24 run, version 1.4.2, sent 534 public and 308 sealed decisions to each system, one request at a time from a server in Germany, and weighted four axes equally: intelligence, calibration, speed and cost. It counts a system as Jev-class when it costs at most twice what Jev does per decision and answers within twice Jev's median latency. Two cautions travel with every score below: the site labels itself a work in progress, and several entrants disclose tuning against the public items.

Official JevBench v1.4.2 scores, published September 24, 2026. decider-4b v2.1 shipped the same day and is not yet ranked. Hardware and license notes come from each project's own documentation.

Which one fits your hardware

No GPU: Laya, a 421-million-parameter encoder that JevBench ran on a CPU, is the one model here that suits an always-on mini PC. Read its own card first: its base checkpoints score near chance on typed-decision workflows until you fine-tune them, and each question is capped at 512 tokens.

8 to 12 GB: the 4B class. decider-4b, lev and Tev1 all build on Qwen3.5-4B, about 8 GB of 16-bit weights, and lev's authors note that a CPU takes seconds per call. The 12 GB RTX 3060 class remains the on-ramp in our local AI models by VRAM guide.

16 GB: Winnow-12B. Its 8-bit GGUF file is 12.67 GB, and its author tested it with a 64K context on a 16 GB RTX 5070 Ti.

48 GB: Cygnet as published. Its recipe runs Gemma 4 12B at full precision and was measured only on 48 GB workstation cards. For the GPU side of that math, see our local AI hardware guide.

Can you use Jev with Ollama?

Not natively. Ollama has no decision-model mode, and a feature request asking it to support System 1 models such as Kev and Laya, opened September 22, is still open with no linked code. Four routes work today:

1. ollaya, the closest thing to Ollama for decision models. This new Apache 2.0 runner uses Ollama-style commands (pull, run, serve) and serves Laya and decider's 2B and 0.8B builds behind TypeSafe's request format; the official SDK works once its base URL points at localhost port 11435. It fetches weights from each author's repository, pinned and checksummed. Its installer pipes a script to your shell, so read the script first or use its Docker image.
2. The model's own server. decider, lev and Laya each ship a server that accepts TypeSafe's /v1/systemone request, so application code does not change. Check the bind address: Laya's server listens on every interface with no authentication unless you set LAYA_API_KEY, and lev's quickstart uses 0.0.0.0 too. Bind to 127.0.0.1 or set a key.
3. Winnow-12B through llama.cpp. Hugging Face offers a one-line Ollama command for Winnow, but that gets you chat only. The typed endpoint comes from Winnow's own llama.cpp-based server.
4. Any instruct model you already run, plus a shim. Ollama returns token probabilities, so small projects such as local-jev and chakuho expose /v1/systemone on top of it. Use an instruct model rather than a thinking model: through Ollama's OpenAI-compatible endpoint, a thinking model's first token starts its reasoning instead of answering. Without a calibration step, the numbers are normalized scores, not calibrated probabilities.

Build it yourself: the launch-day pattern, now measured

The do-it-yourself method from our September 15 pass, once calibrated, now scores within 1.5 points of Jev. That method is one forward pass, the answer restricted to the option letters, and probabilities read from the model's own distribution. Cygnet applies it to a frozen Gemma 4 12B with one fitted temperature and ranks fourth on JevBench at 61.8. A raw Qwen3 4B read the same way without calibration scores 29.1 on calibration, against Cygnet's 74.9. The fitted temperature is most of the difference.

Nokia's applied research group packaged that calibration step as AnyJev, an Apache 2.0 library that fits a closed-form head per question from 100 to 300 labeled examples, with no gradient training. In Nokia's own test, accuracy moved six points, but the share of traffic that could be automated at its chosen risk level rose from 7.7 to 52.0 percent. It reads at most 26 options per question.

On llama.cpp's server the core request is unchanged: one token, its probability distribution, and a grammar that admits only your options. The parameters are documented in the llama.cpp server README.

{
  "prompt": "STATE: (customer message here)\nQUESTION: How urgent is this request?\nOPTIONS: low, medium, high\nAnswer:",
  "n_predict": 1,
  "n_probs": 5,
  "post_sampling_probs": true,
  "temperature": -1,
  "grammar": "root ::= \" low\" | \" medium\" | \" high\"",
  "cache_prompt": true
}

Set n_probs at least as large as your option count, and measure latency on your own hardware with timings_per_token before trusting anyone's number, ours included.

Where Jev still wins, and what none of them can do

Jev keeps the lead on raw judgment, and every model in this class, Jev included, is weak on hard decisions. On JevBench's intelligence axis Jev scores 53.1 to decider-4b v2's 49.4, and it tops the capability ranking at 64.7; decider-4b v2 wins the official composite because it is faster and cheaper to run. Calibration is not unique to TypeSafe's training method, either: Hopper, an open entry, scores 79.1 on calibration against Jev's 76.3.

The sealed set is the sobering part. On JevBench's 308 private decisions, drawn from harder families such as multi-hop reasoning, trade-offs and adversarial traps, Jev answers 36.7 percent correctly and decider-4b v2 34.7 percent. GPT-6 Luna at medium reasoning gets 95.5 percent, at 3.4 times Jev's cost per decision. System One models are a fast, cheap tier for gut-check decisions, not a replacement for a reasoning model.

TypeSafe says much the same on its limitations page for Jev 1.13: the model reads dates as text rather than ordered quantities, and state written to steer it can move the answer, because it does not treat input as hostile by default. Every open clone reading the same kind of text shares that weakness. A decision model used as a guard belongs beside deterministic checks, never in place of them, and on its own segment of your network, as our local AI security guide lays out.

Last, pin what you run by repository and commit. JevBench alone lists at least four unrelated projects named some variant of OpenJev, and two unrelated models are called lev: Interfaze's 4B adapter and Franck Verrot's 350M model, which JevBench ranks 44th.

Who should switch

Home Assistant and homelab builders should start local. TypeSafe's own smart-home demo maps onto a local stack: ask the fan-out questions (request type, room, device, action) of a local decision model, branch in code, and hand anything conversational to a local chat model. Our OpenClaw and Home Assistant guide covers the agent wiring.

Teams pushing millions of routine decisions may find Jev's capability edge worth the dependency. Together AI hosts Tev1 at the same $0.042 per million input tokens, but hosted open weights still run on someone else's server. Whether an open-weight model ever matches Jev on the hard set is the part of this story that decides who controls the infrastructure.

Frequently Asked Questions

Can you run Jev locally?

Not TypeSafe's Jev, which is a closed, hosted API with no download. You can run an open Jev-class model instead. On JevBench's September 24 run the top-ranked system overall was decider-4b v2, an Apache 2.0 model with 4.2 billion parameters, and several open models accept TypeSafe's request format, so the official SDK works against them by changing its base URL.

Is Jev open source or open weights?

No. TypeSafe has published no weights, no on-premises build and no timeline for either. The open models called Jev-like or Jev-class are independent projects, not affiliated with TypeSafe, and each carries its own license: decider, Laya and Winnow-12B are Apache 2.0, while Together AI says the license for Tev1's fine-tuned weights is still being finalized.

Can I use Jev with Ollama?

Not natively. Ollama has no decision-model mode, and a feature request to add System 1 models such as Kev and Laya is still open. The workarounds are ollaya, a community runner with Ollama-style commands that serves Laya and decider, or a small shim that reads answer-letter probabilities from an instruct model Ollama already runs.

What is the best open-source alternative to Jev?

On JevBench's official score for September 24, 2026, decider-4b v2 from Mapika ranks first, 0.8 points ahead of Jev, mainly because it is faster and cheaper to run. Jev still leads on raw capability. For a CPU-only machine, Laya is the practical choice, though it needs fine-tuning to be accurate on most workflows. Rankings change almost daily.

Can Jev hallucinate?

It cannot return an answer outside the options you define, which is what TypeSafe means by "can't hallucinate." It can still pick the wrong option. TypeSafe's launch post says its zero-percent figure is not an empirical measurement, and its limitations page for Jev 1.13 lists weak spots including dates, arithmetic and adversarial input.

What is RLCD?

RLCD, Reinforcement Learning for Calibrated Decisions, is TypeSafe's training method for making Jev's probabilities match how often it is right. Independent measurement now exists: JevBench scores Jev's calibration at 76.3 out of 100, near the top of its class, though Hopper, an open model, scores 79.1, and Laya's authors say they train with a method they also call RLCD.

What is JevBench?

JevBench is an independent benchmark for Jev-class decision models, run by Benchmark Heaven and not affiliated with TypeSafe. Version 1.4.2, scored September 24, 2026, tested 93 systems on 534 public and 308 sealed decisions, weighting intelligence, calibration, speed and cost equally. The site labels itself a work in progress, and its harness and public tasks are MIT-licensed.