YUBA-XV

A stealth checkpoint from Compiwer AI.

Download it, run it, try to break it. We'd like your honest notes: what it gets right, where it fails, what surprised you.

The full reveal (its real name, how it was made, and how it scores on the tests we run) is on Saturday 28 November 2026 at Compile '26, Compiwer AI's developer event: https://www.compiwer.ai/compile

What we can say now

  • It is a fine-tuned open-weights model with about 9 billion parameters, released under Apache-2.0.
  • It thinks before it answers, reads images, and writes Moroccan Darija as well as English, French and Arabic.
  • It was trained with a gate: a training round is kept only if no test score we track drops by more than a small margin we set in advance.

Run it

It loads with Transformers like the open model family it comes from:

from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "CompiwerAI/YUBA-XV"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="auto", device_map="auto")

vLLM: vllm serve CompiwerAI/YUBA-XV

Feedback

Reply to the announcement on X (@compiwer_ai) or email hello@compiwer.ai.

License

Apache-2.0. The licence and notices of the model it was fine-tuned from are included in this repository.

Built by Compiwer AI, Morocco. Ayham Habbad, CEO and Founder.

Downloads last month
15
Safetensors
Model size
9B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support