Julia-1 vs Jev: Local vs Cloud Decision Models (2026)

What Is Julia-1?

Julia-1 is a 144.3M-parameter decision model built by Supersonic Labs on top of mmBERT-small, a multilingual encoder from Johns Hopkins' Center for Language and Speech Processing (CLSP). Given a context, a question, and 2 to 20 candidate answers, Julia-1 scores every option and returns the best one — it does not generate free text.

The model keeps mmBERT-small's attention layout: every third layer attends globally across the full input, and the other layers attend only within a ±64-token local window. A 256,000-row vocabulary embedding accounts for 98M of the 144M total parameters, and a single request only reads the embedding rows for the tokens it actually uses — one reason the model runs on CPU without a GPU.

Julia-1 ships under the Apache 2.0 license, weighs 550.5 MiB on disk (FP32 weights), and accepts up to 8,192 combined tokens across context, question, and candidate answers. It targets classification, routing decisions, ordered scoring, and yes/no judgments — the kind of fixed-choice decision an AI agent makes before calling a tool or escalating a ticket, not open-ended generation.

Its base model, mmBERT-small, is separately notable: JHU CLSP trained it on more than 1,800 languages (1,833 in the final training phase) under the MIT license. That figure describes mmBERT-small's pretraining coverage — Julia-1's own accuracy has only been measured across 52 locales in the MASSIVE benchmark, so treat the 1,800-language number as a base-model fact, not a claim about Julia-1's multilingual accuracy.

How Accurate Is Julia-1?

Julia-1 scores 94/100 on AG News (4-label topic classification) and 86/100 on the DAIR Emotion benchmark (6 labels), both vendor-reported by Supersonic Labs. On the 52-locale MASSIVE intent benchmark it answers 110,573 of 154,648 items correctly (71.5%), and on a mixed set of typed decisions it scores 1,463 out of 2,000 (73.15%).

The weaker result is a 100-example pilot on Banking77, a 72-label banking-intent benchmark: Julia-1 scored 64/100 against an 87% reference score. Supersonic Labs' own model card is explicit that "100-example pilots are encouraging signals, not guarantees," and flags long, ambiguous label sets like Banking77's 72 categories as Julia-1's current weak point.

None of these numbers have been independently reproduced outside Supersonic Labs' own model card as of publication — treat them as vendor-reported until a third party verifies them.

Should You Run Julia-1 Locally or Call Jev's Cloud API?

Jev, launched by TypeSafe AI on September 15, 2026, is the cloud counterpart to Julia-1's local approach — and it launched to far more attention. The launch video drew roughly 40 million views on X, TypeSafe AI raised a $40 million seed round led by DCVC around the announcement, and investors have since approached the company at valuations above $10 billion, per Bloomberg's coverage.

Jev is built by a former OpenAI researcher and is positioned as a "System 1" model: it returns typed, calibrated decisions — not text — for routing, scoring, triage, and moderation, the same task category as Julia-1. Unlike Julia-1, Jev is cloud-API-only through partners such as DigitalOcean's Model Catalog. TypeSafe AI has not published Jev's parameter count or model architecture; outside observers only suspect, unconfirmed, that it is transformer-based on an open-weight foundation.

Pricing is per-token rather than free-after-download: Jev charges $0.042 per million input tokens with output tokens currently free, and TypeSafe AI reports 70–500 millisecond end-to-end latency under standard cloud API conditions. No public benchmark has tested Jev and Julia-1 head-to-head — the two have not been evaluated on the same task set.

For CPU-only local setups outside the decision-model category, see best Ollama models for CPU-only machines.