Tencent Hunyuan

Tencent

2026-09-22 07:22:42

Tencent rolls out Hy Image3.5 preview, says human evaluation win rate is 30% above Hy Image3.0

Tencent Hunyuan said on Sept. 22 via its official X account that the preview version of its image generation model, Hy Image3.5, is now live. The company said the new model posted a 30% higher win rate than Hy Image3.0 in human evaluations. It supports both text-to-image and image-to-image generation, offers up to 2K resolution, and improves image consistency. Tencent has also listed the preview model on Tencent Cloud API, where each generated image costs $0.024. The company said only generated images are billed, while uploaded reference images are free of charge. Tencent is also offering a two-week free trial, though access is limited to the Miora and OnSolo platforms. In its announcement, the team invited users to try the model and report failures, writing: 「试试看，告诉我们哪里会出错」.

330

Tencent Hunyu

2026-09-20 08:58:39

Tencent Hunyuan, Tsinghua and Peking University launch IWC-Bench to test whether AI-built webpages actually work

Tencent Hunyuan, together with Tsinghua University and Peking University, has introduced IWC-Bench, a benchmark designed to evaluate how usable AI-generated webpages are in real-world use. Instead of relying mainly on code inspection or screenshots, the benchmark has an agent open the webpage, click buttons, enter text, switch pages, and then verify which functions actually run. The system scores three areas: visual quality, ease of use, and whether the requested user functions were delivered. According to the release, IWC-Bench includes 369 real user requests and 5,088 acceptance criteria, while 17 models generated a total of 6,273 web applications for testing. In a separate comparison covering 197 manually reviewed cases, IWC-Bench matched human choices in 168 cases, for an agreement rate of 85.3%. The results also suggest that a page looking polished does not necessarily mean it is usable. GPT-5.6-Sol ranked first in visual appeal, but only sixth in usability. For individual webpages, the correlation between appearance and usability was 0.36.

420

Yao Xingcheng

2026-09-15 09:58:15

Former Tencent Hunyuan pre-training lead Yao Xingcheng joins Thinking Machines Lab

Yao Xingcheng, the former head of pre-training for Tencent’s Hunyuan large language model, has joined Thinking Machines Lab, according to a BlockBeats flash report. The startup was founded by former OpenAI CTO Mira Murati. Yao left Tencent in July this year, and there had previously been market talk that he might join Meta. Caijing-style outlet Leifeng, citing a person familiar with the matter, said Thinking Machines offered Yao compensation that was clearly higher than what he received at Tencent. The report added that his annual pay package at Tencent had already reached tens of millions of yuan. Yao graduated from Tsinghua University’s Yao Class. Before joining Tencent, he worked on large model research at Moonshot AI and was involved in the development of Kimi K2 and Kimi Linear. He is also listed as co-first author of the widely known sentence embedding method SimCSE.

600

Tencent Hunyu

2026-09-15 09:11:00

EvolveScaler tests whether LLMs can keep up with changing information, with the best GPT-5.5 score at 59.3%

Teams including Tencent Hunyuan have released EvolveScaler, a benchmark built to test whether large language models can reason over information that keeps changing over time. The setup injects updates, reversals, additions, and invalidations into long passages, then asks models to answer based on the latest state rather than retrieve a ready-made sentence from the text. One example starts with 40 days of game records and then asks a counterfactual question about what would happen if a small boss had not been defeated on day 7, forcing the model to recompute later equipment, health, and battle outcomes. The benchmark includes 117 task categories and 159 question types, with the longest sequences containing about 1,200 events. It evaluates 14 model configurations, including GPT-5.5, Gemini 3.1 Pro, DeepSeek V4 Preview Pro, GLM-5.2, and Qwen3.5 Plus. In the hardest setting, the median score was only 11.3%, while the top-performing GPT-5.5-xhigh reached 59.3%. The team also said the dataset can be used for training: after adding 6,000 such samples, an internal A3B model improved on all eight external tests, gaining an average of 5.25 points.

800

Tencent Hunyu

2026-09-10 12:14:24

Tencent open-sources AuK, a 1.5B speech model for voice editing and cloning

Tencent Hunyuan has open-sourced AuK, a 1.5 billion-parameter speech generation and editing model that the company describes as an "audio version of Nano Banana." The model combines voice cloning, word-level correction, emotion transfer, accent removal, speaking rate and pitch adjustment, denoising, and multi-speaker and music separation in a single system, with natural-language controls. According to the announcement cited by BlockBeats, AuK can edit existing recordings directly. Users can change only the incorrect words in a clip instead of re-recording the full segment. It can also change emotion, timbre, or accent without altering the spoken content, and supports tasks such as lyric editing, removing laughter or coughs, and turning normal speech into a whisper. Tencent said AuK led several speech generation and general speech editing benchmarks. On SpeechEditBench, it scored 49.73, compared with 28.70 for the runner-up, Ming-UniAudio. Tencent also released AuK-Flash, a distilled version that needs only four inference steps and runs about 4.5 times faster. The paper notes that AuK still cannot reliably understand all free-form natural-language instructions and currently depends on a Prompt Enhancer to classify tasks, organize parameters, and rewrite prompts. The code and model weights have been released under the MIT license.

830

AI Models

2026-09-07 02:55:04

Global AI model usage hit 115 trillion tokens last week, with Chinese models taking four of the top five spots

Global AI large-model usage reached 115 trillion tokens in the week from Aug. 31 to Sept. 6, according to calculations by National Business Daily based on the latest OpenRouter data. That marked a 1.77% increase from the previous week. China accounted for 56.72 trillion tokens, up 2.83% week over week, and stayed ahead of the United States for a 19th straight week. U.S. model usage came in at 16.54 trillion tokens, down 3.1% from the prior week. Among the five most-used models globally, four were Chinese. Tencent Hunyuan Hy4 preview ranked first with 14.7 trillion tokens, up 379% week over week. GPT-5.6 Luna placed second with 12.9 trillion tokens, while Zhipu GLM-5.3 Flash rose to third with 12.4 trillion tokens. DeepSeek-V4-Flash official and preview versions ranked fourth and fifth. MiniMax M3 returned to the ranking after nearly a month, taking sixth place, and has become the foundation model for the Arabic large model of Saudi Arabia’s HUMAIN. Xiaomi MiMo-V2.5 and Gemini 3.7 Flash dropped out of the ranking this week.

960

AI models

2026-09-07 03:04:13

Chinese AI models lead global usage for a 19th straight week, with Tencent's Hy4 preview at No. 1

Chinese AI models remained ahead of the U.S. in weekly large-model usage for the 19th consecutive week, according to calculations published by National Business Daily based on the latest OpenRouter data. For the week of Aug. 31 to Sept. 6, global AI model calls reached 115 trillion tokens, up 1.77% from the previous week. China accounted for 56.72 trillion tokens, a 2.83% weekly increase, while the U.S. logged 16.54 trillion tokens, down 3.1%. Four of the top five models by usage were Chinese. Tencent Hunyuan Hy4 preview ranked first with 14.7 trillion tokens, posting a 379% week-over-week increase. GPT-5.6 Luna came in second at 12.9 trillion tokens, and Zhipu's GLM-5.3 Flash rose to third with 12.4 trillion tokens. DeepSeek-V4-Flash, in its official and preview versions, placed fourth and fifth. MiniMax M3 returned to the ranking after nearly a month and took sixth place. The report also said the model has become the foundation for HUMAIN's Arabic large language model in Saudi Arabia. Xiaomi's MiMo-V2.5 and Gemini 3.7 Flash fell out of the ranking during the week.

960

Tencent Hunyu

2026-09-02 09:36:15

Tencent Hunyuan Hy4 Preview Goes Live on B.AI API

On September 2, the B.AI platform announced that Hy4 Preview, the latest open-source flagship model from Tencent's Hunyuan team, has officially launched through its API service and is now fully open for developer access. Hy4 Preview is described as an early version in the Hy4 series. The underlying architecture is a 770B sparse mixture-of-experts (MoE) design, with 49B parameters activated for every token processed. The model also provides an extended context window of up to 1 million tokens. According to B.AI, that combination gives the model professional-grade performance in software engineering agents, office data workflows, game development, and scientific research. Starting immediately, developers can call Hy4 Preview directly through the official B.AI API channel. For further information, they can visit chat.b.ai/chat for more details and immediate testing. B.AI also said that going forward it will continue to introduce frontier models, providing developers with a richer and more cost-effective range of AI compute options.

1130