On July 6, Tencent officially released the 295B parameter version of the Hy3.0 (Hybrid Expert Model) MoE. The most significant change is the switch to the permissive Apache 2.0 open-source license, which removes the previous regional restrictions that prohibited use in the EU, UK, and South Korea. When the preview version was released in April, many overseas or multinational teams were deterred by strict regional limitations and the commercial threshold of 100 million monthly active users. The official version has eliminated all compliance barriers and made substantial optimizations to address deployment challenges: it integrates fast and slow thinking mechanisms and incorporates a 3.8B parameter Multi-Token Prediction (MTP) layer for parallel generation, effectively reducing inference latency. Through fine-grained data cleaning and training constraints, the official version has reduced the hallucination rate of the large model from 12.5% to 5.4%, and the error rate in multi-turn interaction tests has dropped from 17.4% to 7.9%. To address the persistent issue of tool invocation inaccuracies during agent development, Hy3.0 has also undergone targeted enhancements. Whether in mainstream frameworks like Cline or CodeBuddy, the accuracy fluctuations of cross-framework tool calls have been kept under 4%, resulting in more stable outputs. The officially released FP8 quantization version has also lowered the memory requirements for local deployment and fine-tuning.
All Comments