Zubr: ZuBrIQ models go open — our build on top of open source

On July 13, 2026, the ArtCloud team published the Zubr model family on Hugging Face — language models of the ZuBrIQ platform that run inside the Republic of Belarus.

First, a note on what stands behind the name. Zubr is not a set of models written from scratch. The foundation is proven open models from the Qwen family, published by the community. ArtCloud took them and did the rest of the work that turns an open source model into a working service: quantization and a production-ready build for inference, configuration for serving in Belarus, the platform-side integration, and testing on Russian- and Belarusian-language tasks. By publishing the models openly, we give back to the community what we ourselves received from open source.

The first release included two models:

- Zubr1.0-35B-A3B-AWQ — the flagship. The open Qwen3.6-35B-A3B MoE model (35B parameters, ~3B active per token) in AWQ quantization (~24 GB of weights) with a context of 131,072 tokens — ready for inference on vLLM, including tool calling.
- Zubr1.0-VL-4B — a vision model based on the open Qwen3-VL-4B: understands images, scans and screenshots; powers multimodal requests and the Vision Bridge (screen understanding) in ZuBrIQ.

Legal, local, OpenAI-compatible

The main point of the release is legal clarity. Both models are distributed under the Apache 2.0 license and run on infrastructure in Belarus: no request leaves the country, billing is in Belarusian rubles. The flagship is served through an OpenAI-compatible API at `api.zubriq.by` (alias `zubr`): existing code built on the OpenAI SDK switches to Zubr by changing a single line of configuration.

For companies this is a practical alternative to the "either foreign cloud AI or nothing" choice: chat, document work, tool calling and multimodality — legal, without leaving the country.

An open ecosystem

The models and weights are published on Hugging Face — download them, run them on your own hardware, or adapt them to your task. The base Qwen models are open, and so are our builds: companies can inspect what is inside and assemble their own solution on this foundation.

The family keeps growing: after the July release, ArtCloud trained and published Zubr1.5-VL-32B — the first fine-tuned model of the family. The team attached a QLoRA adapter to the open Qwen3-VL-32B and trained it on the legislation of the Republic of Belarus (28 codes and their application practice); the model understands document scans and handles legal scenarios.

The Zubr models are part of the ZuBrIQ platform, running since August 2025: LLM chat, document work, assistants and BYN billing at zubriq.by.