UHR-BAT2
Released by RL-MIND, 2026.

This model is derived from Qwen/Qwen3-VL-8B-Instruct, distributed by
the Qwen team under the Apache License, Version 2.0.
Upstream: https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct
Architecture and software: https://github.com/QwenLM/Qwen3-VL

Modifications: the multimodal merger and language model were fine-tuned
on a selected 10,000-example GeoLLaVA training manifest. The vision
encoder was frozen. Full fine-tuned weights are distributed, rather
than a LoRA adapter. The native Qwen3-VL architecture is retained.

This release adds a model card, inference example, aggregate evaluation
records and release verification metadata. Benchmark and training data
are not included. Their respective terms remain applicable.

The upstream revision used during training was not recorded. A current
upstream revision is not represented as the training revision.
This derivative release does not imply endorsement by the Qwen team.