A System One model.

System One doesn’t need the whole LLM.

A language model is built to write; its deepest layers mostly shape the next word. A System One model never writes. It reads, then decides. So v5.0-VL keeps the first 20 of Qwen3.5-4B-Base’s 32 layers: 3.25B parameters, one pass.

Trained once per exit layer with the same recipe, the cut at 20 scores like all 32 on the decision suite (0.760 vs 0.761) and the held-out set (0.688 vs 0.686). Knowledge-heavy MMLU-Pro is the exception: 0.422 vs 0.457.

Show it a picture.

It picks the answer, with a confidence. Send one to four images with a request: a photo, a screenshot, a chart. Ask the same typed questions as with text.

Same picture, same question.

RSI-Jev and a chat model.

Replay of measured timings on one GB10

RSI-Jev v4.0-VL2B, one pass0 ms

Qwen3.5-2B chat2B, writes its answer as JSON0 ms

Race again

All 30 questionsRSI-JevChat model Median time69 ms540 ms Valid answers30 / 3020 / 30 Probability per answeryesno, text only

RSI-Jev’s time is its median over 9 calls; the chat model’s text appears at the token times of one streamed call (greedy, up to 256 tokens). Which one is more accurate is unresolved: most of the chat model’s invalid replies gave an answer under the wrong JSON key; read leniently, it gets 23 of 30 right to our 26, a gap 30 questions cannot settle. Leaves photo by Wald1siedel (CC BY-SA 3.0; the resized copy is shared under CC BY-SA 3.0). The receipt is generated; its blank lower half is cropped here.

A probability for every answer.

One leaking capsule in the tray, and it has never seen a defect example

Production-line inspection photo of green gel capsules.

Is everything in the photo good, or is at least one part defective?

68%Defective

• Defective68%
• All good32%

✓ matches the reference254 ms on one GB10

CapsulesBoardRed lightSpeed limitTeddy bearChartDialogCheckout

Real output from v4.0-VL on one GB10; the time is the whole call. The three Wikimedia photos are shown at higher resolution than the 512 px copies the model was sent. Photos: VisA (Zou et al. 2022, CC BY 4.0, resized); Wikimedia Commons speed sign by P.Ctnt (public domain) and traffic light by Jacklee (CC BY-SA 3.0; the resized red-light photo is shared under CC BY-SA 3.0) and teddy bear via Pixabay (CC0). The charts and screens are generated by us.

Finding defects on factory parts, with no defect examples.

Same API as TypeSafe’s Jev, so existing Jev code works. Not affiliated with TypeSafe.

How the agents build it

The chart runs to v3.0. v4.0-VL then added two rounds of image training and kept the text score at 0.756. Its last stage is RL that penalises confident mistakes: held-out calibration error before any post-hoc fix fell from 0.200 to 0.082 (one seed). v5.0-VL cuts a 4B model at layer 20 and retrains its decision head to read each option from its own tokens: Decision Index 38.38.

The experiments were run by the next generation of AutoScientists, our AI research-agent system. Read the experiment log: every arm through v3.0, and every arm on the v4.0-VL and v5.0-VL lines.

You ask. It picks the answer, with a confidence.

Real output, in about 35 ms on a desk-side GPU. It never writes free text, so there is nothing to parse and nothing made up. (Measured on v3.0.)

Three commands

pip install "rsi-jev[fast,vision] @ git+https://github.com/Shanghua-Gao/RSI-Jev"

rsi-jev serve v5.0-vl-3b

curl localhost:8000/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": [{"role": "user", "content": "I was charged twice. Please refund."}],
  "questions": {"refund": {"type": "noul",
    "instructions": "Does the user request a refund?"}}}'

Already using Jev? Point your client at http://localhost:8000 instead. For an image, add an images list; the README shows how.

More examples

Recorded output, measured on v3.0.

How long does an answer take?

The support ticket above: one request, three questions. Median of several runs (measured on v3.0).

Jev-style models answer in one step instead of writing text, which is why they are all this fast. Jev’s time includes the trip over the internet from our machine. The chat model is Qwen3.5-2B, the same size as RSI-Jev, writing its answer as JSON.

And faster than when v3.0 first shipped

We kept working on how it runs. The same model on the same machine now answers 1.5–5.8× faster, with the same answers.

How we made it faster, and what didn’t work

Built and measured on an HP ZGX Nano. Thanks to HP and NVIDIA for providing the HP ZGX Nano AI Station, powered by the NVIDIA GB10 Grace Blackwell Superchip.