Sansar OCR 2B
Copyright 2026 Muse Mesh Private Limited

This product is a modified version of Qwen3.5-2B (https://huggingface.co/Qwen/Qwen3.5-2B, revision
15852e8c16360a2fea060d615a32b45270f8a8fc), Copyright 2026 Alibaba Cloud, licensed under the Apache License,
Version 2.0 (LICENSE-QWEN).

Modifications by Muse Mesh Private Limited (2026-10): all weights (language model and vision encoder) fine-tuned on
Sanskrit page-OCR data; weights stored in bfloat16 except the parameters the base release keeps in float32
(A_log, linear-attention norms); config.json dtype/use_cache and generation_config.json (stop tokens, greedy
decoding) set for inference. The tokenizer, chat template and processor settings are the base model's.