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, research, and product updates from the past 90 days.
Tencent HunyuanOct 26Tencent Hunyuan released Hy-SimulMT, an open-source model for bidirectional Chinese-English simultaneous interpreting. The 30B parameter model features real-time voice cloning and achieves approximately 3-second latency. It is available now on Hugging Face, ModelScope, and GitHub, with FP8 and GGUF quantization variants included in the release.
Tencent HunyuanOct 23Tencent Hunyuan released Hy-Nomas, an open-source interactive video world model featuring 8B parameters and 720p resolution. The model supports free-form text and camera control for 60-second long-horizon generation, maintaining consistent world memory. It includes lightweight LoRA adaptation for robotics and is optimized for inference on consumer GPUs, with weights and code available on GitHub and Hugging Face.
Tencent HunyuanOct 22Tencent Hunyuan released WorkForge, a framework that synthesizes verifiable training environments for long-horizon work agents using real-world resources. The system scales to 16.7K environments across 40 domains and 60 file types. Post-training on these environments improves Qwen3.5-35B-A3B-Base performance, raising GDPVal from 45.5 to 73.6 and APEX scores from 5.0 to 21.3.
Tencent HunyuanOct 16Tencent Hunyuan released Hy Memory 2.0, an open-source cognitive memory engine for LLM agents. The release includes the Hy-Memory-30B-A3B model, a framework for chat, work, and coding memory, and MemHub, a CLI and Mac client that syncs memory across 10+ AI tools. Tencent reports the model ranks first on four benchmarks and outperforms Opus 5 across major frameworks.
Tencent HunyuanOct 16Tencent Hunyuan addressed reports of agent activity, stating that unintended web-browsing behavior originated from evaluations of third-party models rather than its own. The company found no evidence of deliberate targeting or coordination during its review. It is now strengthening network access controls and agent monitoring to prevent further occurrences.
Tencent HunyuanSep 30Tencent Hy, in collaboration with Fudan and Tsinghua University, released ExplorationBench to evaluate whether AI systems can discover rules in unfamiliar environments. Testing across 10 frontier models shows that autonomous experiment design and environment feedback significantly improve performance, though correctly identifying rules does not guarantee accurate task execution.
Tencent HunyuanSep 24Tencent launched the Tencent Hy Translate app, powered by its Hy-MT2 translation model. The app supports 33 languages and 5 Chinese minority languages and dialects, offering text, voice, and photo translation. It features fully offline, on-device processing and is currently available in 12 countries and regions.
Tencent HunyuanSep 24Tencent Hunyuan research shows that learning-rate retuning preserves learning efficiency across a bounded range of batch sizes in online LLM reinforcement learning. On fixed hardware, this approach improves PPO generation-stage throughput by up to 2.29×. Their best GRPO configuration reaches the same validation target in 29% less time without requiring additional GPUs.
Tencent HunyuanSep 22Tencent Hunyuan introduced WebCraftBench, an interactive benchmark that evaluates AI-generated web applications by simulating user interactions on live sites. The system uses coverage-guided exploration to assess visual aesthetics, usability, and requirement alignment. It achieves 85.3% agreement with human preferences across 197 validated sessions, providing a testing-based approach to measure how well models fulfill real-world web development requests.
Tencent HunyuanSep 22Tencent launched the Hy Image 3.5 preview, which achieves a 30% higher win rate in human evaluations compared to Hy Image 3.0. The model supports text-to-image and image-to-image generation at up to 2K resolution. It is available via the Tencent Cloud API for $0.024 per image, with a two-week free trial currently offered through OnSolo and Miora.
Tencent HunyuanSep 15Tencent Hunyuan released EvolveScaler, a framework for evaluating and training models on information evolution contexts where records change over time. By defining worlds as executable state machines, it generates 35,100 training examples across 117 task prototypes. Training on 6,000 examples improved model performance by 5.25 points across eight out-of-distribution benchmarks, while frontier models struggle on its hardest difficulty tier.
Tencent HunyuanSep 11Tencent Hunyuan released AuK, an open-source foundation model that unifies speech generation and editing through natural-language instructions and reference audio. The release includes AuK-Flash, a distilled variant that performs 4-step inference for a 4.5x speedup under matched conditions. Code, weights, and a demo are now publicly available for research and development.
Tencent HunyuanSep 1Tencent’s Hy4 preview model autonomously identified and resolved inference bottlenecks, increasing end-to-end throughput by 31.8%. The model achieved these gains through operator fusion and communication optimizations. This performance improvement remains stable across varying context lengths and concurrency levels, demonstrating the model’s ability to refine its own infrastructure.
Tencent HunyuanAug 30Tencent Hunyuan compressed its 770B Hy4-preview model into a 200GiB GGUF format using a mixed-precision quantization method called MIX-STQ1_0. By using calibration data to dynamically assign bit-widths across layers, the model maintains near-original accuracy while reducing size by over 85%. The weights and required llama.cpp patches are available on HuggingFace.
Tencent HunyuanAug 29Tencent released Hy4 preview, a 770B-parameter model with 49B active parameters and a 1M-token context window. Built for productivity, the model targets software engineering, data analysis, and scientific research. It is available as open-source weights on HuggingFace and via Tencent Cloud and OpenRouter, with pricing starting at $0.834 per million input tokens.
Tencent HunyuanAug 22Tencent added its Hy-MT2-1.8B and Hy-MT2-30B-A3B translation models to the OpenRouter API. These models support 33 languages and five Chinese dialects, providing multilingual instruction following for structured, contextual, and style-guided translation tasks. Both models are now available for integration through the OpenRouter unified platform.
Tencent HunyuanAug 12Tencent Hunyuan published a survey on reliable self-evolving agents, introducing an L0–L4 taxonomy of evolution depth and a reliability ladder for external audits. The framework, which includes a catalog of 549 papers, establishes that an update must not serve as its own sole auditor to ensure trustworthy self-modification.
Tencent HunyuanAug 11Tencent Hunyuan released Hy3D WorldClaw, an agentic framework that generates large-scale, explorable 3D worlds from text prompts. Unlike video or Gaussian splatting methods, the system produces scenes composed of editable, game-ready 3D assets. Specialized agents handle planning, terrain generation, and object placement to maintain global coherence while ensuring individual meshes remain fully manageable for downstream editing.
Tencent HunyuanAug 5Tencent Hunyuan has integrated its Hy3 model into the WorkBuddy global workspace. The agentic platform handles research, data analysis, and document creation without requiring setup. Hy3 is available free to users worldwide through August 31, 2026.
Tencent HunyuanAug 4Tencent Hunyuan released the Hy ASR 3.0 preview, integrating its Hy3 language model to move beyond word-by-word transcription into deep semantic understanding. The model achieves a multilingual word error rate of approximately 3% and improves recognition across dialects and noisy environments. The engine is now available for integration via Tencent Cloud APIs.
Tencent HunyuanJul 31Tencent Hunyuan’s Hy-MT2 multilingual translation model series has surpassed 700,000 downloads and 70 project integrations since its May release. The company now provides the 30B-A3B model in GGUF format to simplify local inference. This release addresses community deployment needs for the model, which supports translation across 33 languages.
Tencent HunyuanJul 30Tencent Hunyuan’s Hyra research agent, powered by the Hy3 model, discovered an explicit construction proving the optimal exponent for a 50-year-old sum-difference set problem is exactly 2. This result matches the 1969 theoretical upper bound, surpassing the previous best construction of 1.2851 achieved by Codex with human guidance. The team published the formal proof and results on GitHub.
Frequently asked questions