According to Vision One Beating monitoring, Tencent has officially released the Element 3.0 (Hy3) Hybrid Expert Model (MoE) with a 295B parameter setting. The most significant change is the switch of the open-source license to the most permissive Apache 2.0, abolishing the previous regional restrictions that prevented the use in the EU, UK, and Korea.
When the preview version was released in April this year, many international or cross-border teams were discouraged due to strict regional limitations and a 100 million monthly active commercial user threshold. The official version has removed all compliance barriers and addressed the implementation challenges through hardcore optimization: it integrates a fast-and-slow thinking mechanism and incorporates a 3.8B parameter Multi-Token Prediction (MTP) layer for parallel generation, significantly reducing inference latency. Through fine-grained data cleaning and training constraints, the official version has reduced the gibberish illusion rate of large models from 12.5% to 5.4%, and the error rate of multi-round interactive testing from 17.4% to 7.9%.
To address the persistent issue of tool invocation deviation during intelligent agent development, Element 3.0 has also been selectively strengthened. Whether in mainstream scaffolds such as Cline or CodeBuddy, the fluctuation in cross-framework tool invocation accuracy has been kept within 4%, resulting in more stable outputs. The officially synchronized FP8 quantized version has also lowered the local deployment and fine-tuning memory thresholds.
