According to CoinMeta, Baidu released and open-sourced Unlimited-OCR 2, which has approximately 3 billion parameters, but only about 500 million parameters are activated during inference. The second version introduced the sr-swa sparse attention mechanism, making the single-step decoding speed about 5.6 times faster than the old version, and the peak显存 usage decreased from 47GB to 1.77GB. In a 100-page document test, the generation throughput increased from 360 tokens per second to 1315 tokens. Unlimited-OCR 2 scored 77.06 points in Baidu's own benchmark tests, performing better than Gemini 3, Flash, and Qwen 3.8 Max.