Cicero NVFP4
The NVFP4 checkpoint of Cicero, the weights we serve in production. BF16 weights are in butler-compute/Cicero.
Cicero is a decision model: it reads a state (text, a chat, or up to 4 images) and up to 64 typed questions, and returns a calibrated probability for every option of each question. No text is generated: each question is answered in one forward pass instead of a generated answer.
- 71.33 on the public JDI (Jev Decision Index), with 3B active parameters.
- 41 ms median latency.
- Ahead of Decider v1.2 in every area.
- Best on Arts & Human Taste of the models compared below.
- Reads images (up to 4) and contexts up to 131,072 tokens.
Model Details
| Property | Value |
|---|---|
| Base | Qwen3.5-35B-A3B, 3B active |
| Format | NVFP4 experts, 23 GB |
| Questions | yes/no, choice, score; up to 64 per request |
| Output | a probability per option |
| Context | 131,072 tokens |
| Inputs | text, chat, up to 4 images |
How it works
A request is a state and typed questions. We lay each question out as plain text (the context, then the question, then its options lettered A, B, C) and the model reads that prompt in a single forward pass. The answer is read off the model's scores for the option letters at the end of the prompt, turned into probabilities and calibrated for each question type.
| Question type | Asks | Answer |
|---|---|---|
noul |
a yes/no question | probability of yes and no |
choice |
pick one of up to 255 options | a probability per option |
score |
rate on a scale, e.g. 1-5 | a probability per score |
Examples
Two real rows of the public JDI suite, with the probabilities Cicero returned.
BANKING77, one choice question over 77 intents: "I still have not received my new card, I ordered over a week ago."
| Intent | Probability |
|---|---|
| card_arrival | 0.973 |
| card_delivery_estimate | 0.021 |
| card_not_working | 0.001 |
| card_linking | 0.001 |
BFCL, one conversation, five yes/no questions: "Can you find me the recent news articles about Son's latest goal in the sports category?" Question: invoke tool?
| Tool | Probability |
|---|---|
| HNA_NEWS.search | 0.999 |
| HNA_WQA.search | 0.001 |
| OpenWeatherMap.get_current | 0.000 |
| ControlAppliance.execute | 0.000 |
| cookbook.search_recipe | 0.000 |
Performance
Jev Decision Index (JDI) 0.4.1, public suite: chance-corrected skill (0 = chance, 100 = perfect). Cicero's scores are from our own run of the public suite with the official kit. Best in each row in bold.
| Area | Cicero | mammoth-1 | elephant-1 | Decider v1.2 | Drex Max |
|---|---|---|---|---|---|
| Tools & Automation | 87.0 | 88.6 | 88.0 | 85.4 | 81.9 |
| Retrieval & Classification | 76.1 | 77.4 | 76.2 | 68.1 | 68.8 |
| Language Understanding | 74.6 | 76.1 | 73.8 | 71.9 | 72.2 |
| Arts & Human Taste | 64.4 | 64.1 | 61.6 | 56.2 | 53.4 |
| Knowledge & Reasoning | 56.1 | 68.0 | 60.6 | 55.3 | 57.1 |
| Public JDI | 71.33 | 75.34 | 72.24 | 67.74 | 67.52 |
The chart shows the board's decision models that are available on OpenRouter.
Training
Cicero is a post-train of Qwen3.5-35B-A3B, a mixture-of-experts model with 35B parameters and 3B active. The vision tower is the base model's.
Citation
@misc{cicero2026,
title = {Cicero},
author = {{Butler Compute, Inc.}},
year = {2026},
note = {Decision model},
url = {https://butlercompute.com/research}
}
- Downloads last month
- 8