The latest AI news and updates of Tencent Hunyuan — AI division of Tencent building open-source multimodal models for language, image, and video. Covering Tencent Hunyuan's latest launches, product updates, and research from the past 90 days.
Tencent Hunyuan6h agoTencent Hunyuan released AngelSpec, a torch-native training framework for speculative decoding that supports MTP and block-parallel methods. The release includes training code and drafter weights for the Hy3-A21B model. Using the DFly drafter, the model achieves a 1.98–2.40× speedup over autoregressive decoding and 10.5–11.8% higher throughput than DFlash across tested concurrency levels.
Tencent HunyuanJul 21Tencent Hunyuan introduced Hyra-1.0, its first research agent designed to recursively improve solutions for performance-driven research and engineering tasks. The agent features demonstration modules across three domains: AI4AI, AI4Science, and AI4Fun. This release marks the debut of the Hunyuan Research Agent series, focusing on iterative refinement for complex technical workflows.
Tencent HunyuanJul 14Tencent released 1-bit and 4-bit quantized versions of its 295B Hy3 Mixture-of-Experts model. These GGUF-formatted files enable serving the flagship-scale model on a single GPU using llama.cpp. The release includes support for Multi-Token Prediction to improve inference efficiency on consumer hardware.
Tencent HunyuanJul 13Tencent Hunyuan’s Hy3 model now holds the number one position on the OpenRouter LLM leaderboard. The model reached this milestone by processing 6.13 trillion tokens in weekly usage volume. This ranking reflects real-world developer traffic on the platform, marking a significant shift in model adoption.
Tencent HunyuanJun 8Tencent Hunyuan, in collaboration with several institutions, introduced MMAE, the first comprehensive benchmark for instruction-based audio editing. This new evaluation reveals that current AI models achieve an Exact Match Rate (EMR) below 5% overall, plummeting to 0% for complex, mixed-modality tasks, highlighting a significant gap in AI's ability to precisely modify existing audio based on natural language instructions.
Tencent HunyuanJun 7Tencent Hunyuan has officially released Hy-Memory, a memory plugin designed for long-term collaborative AI agents. It uses a 6-layer memory framework and dual System1/System2 processing to enable agents to remember durably and efficiently, reducing memory count by over 70% and token usage by 35% on ultra-long contexts. This aims to move agents beyond single-session context, allowing them to build a persistent, evolving understanding of user preferences and intentions.
Tencent HunyuanJun 7Tencent Hunyuan, in collaboration with Renmin University of China, open-sourced PlanningBench, a framework for evaluating and training large language model (LLM) planning capabilities. This release aims to help LLMs move from generating text to autonomously executing multi-step actions in complex scenarios.
Frequently asked questions