Woofun AI reports that Tencent has officially open-sourced the Element 3.0 (Hy3) Hybrid Expert Model with 295B parameters under the Apache 2.0 license. This update eliminates previous regional restrictions affecting the EU, UK, and Korea, while removing the 100 million monthly active user commercial threshold. The model integrates a fast-and-slow thinking mechanism and a 3.8B parameter Multi-Token Prediction layer to reduce inference latency. Technical optimizations lowered the gibberish illusion rate from 12.5% to 5.4% and multi-round interaction error rates from 17.4% to 7.9%.

Additionally, tool invocation accuracy fluctuations in frameworks like Cline are maintained within 4%, and an FP8 quantized version reduces local deployment memory requirements.