AI Trends

Tencent Hunyuan’s Yao Shunyu Team Answers Nearly 100 Questions: No Deliberate Benchmark Chasing, Larger Models to Be Trained

Sina TechUpdated Nov 22 13:03

AI summary

Tencent Hunyuan’s Yao Shunyu team said it will continue to follow the Scaling Law, with the next-generation Hy model significantly increasing its parameter count, while not deliberately chasing high benchmark scores. The existing Hy4 has 770B total parameters, while Hy4 flash has 224B. The team also plans to launch Hy4-mini and is considering developing a 24B-to-31B dense model and Flash Lite.

Body · AI translation

Added copy for beta-testing promotion Added copy for beta-testing promotionend Add app-exclusive traffic redirection

After releasing Hy4 and Hy4 flash in quick succession in November, Tencent began to gradually reverse perceptions that it had “fallen behind in AI.”

In the early hours of November 21, Tencent Chief AI Scientist and head of the Hunyuan large model Yao Shunyu and his team held a 120-minute Q&A session (AMA, Ask Me Anything) on the technology community Reddit, answering questions about Hunyuan’s R&D direction and approach.

The models’ parameter counts drew attention. In particular, as ByteDance and Kimi move toward models with more than one trillion parameters, why does Hunyuan remain at around 700 billion?

Regarding Hy4, which has 770B total parameters (B denotes one billion parameters), and Hy4 flash, which has 224B parameters, the Hunyuan team said the new models are intended more for general-purpose use than for a strong focus on programming. Hy4 still has considerable room for improvement, especially in post-training, and the team is investing heavily in large-scale reinforcement learning.

The Hunyuan team also said it is planning to train models with larger parameter counts. It believes in the Scaling Law and will continue investing, with the next generation of Hy models set to increase significantly in scale.

Many users were also interested in plans to release models with smaller parameter counts, noting their potential for applications such as on-device use. The Hunyuan team responded that a smaller Hy4-mini is already on the roadmap and that it is considering developing a dense model with 24B to 31B parameters. A cutting-edge Flash Lite is also on the way. The team will also consider creating an even smaller image model and asked the public to keep an eye out for its video model as well.

“Hy3 and Hy4 have already shown that solid smaller models can also outperform larger ones. More importantly, they cost less,” the Hunyuan team said.

One user posted a table outlining the stages of AGI (artificial general intelligence) development, showing that Level 2 represents the median performance of a skilled adult, while Level 3 represents the top 10% of skilled adults. The Hunyuan team responded that Hy4 is currently at Level 2.5.

On the technical front, one user noted that Hy4 sometimes spends too long reasoning. Regarding the issue of “overthinking,” the Hunyuan team said Hy4 aims to maximize intelligence, so excessively long chains of thought (CoT) are a known issue. However, the team distilled it into the more cost-effective Hy4 flash, which uses reinforcement learning to ensure shorter chains of thought. The Hunyuan team also acknowledged that “overthinking” is the behavior most difficult to improve during post-training.

The Hunyuan team also revealed that it is making creative writing and readability key areas for improvement, with room for further progress. It also believes that data and evaluation are the most important tools for achieving AGI: data determines a model’s capabilities, while evaluation guides it in the right direction. The team is also researching continual learning technology—where models learn new tasks from a continuous stream of data while avoiding the loss of old knowledge—which it considers a highly promising direction.

Since Yao Shunyu took charge, the Hunyuan large model team has abandoned the approach of chasing leaderboard positions through benchmark scores. Some users noted that Hunyuan had not published results for common benchmarks such as GPQA Diamond and SWE Bench.

The Hunyuan team responded that these benchmarks were released relatively early and their scores have already become saturated. Instead, the team presents results from newer, more challenging benchmarks, such as Agents'Last Exam. It emphasized that during model development, it addresses certain issues through better training data and diverse internal evaluations rather than deliberately chasing leaderboard rankings.

Some users also asked whether Hunyuan could catch up with frontier models amid restrictions on the supply of advanced AI chips. The Hunyuan team responded that it is currently squeezing the utmost performance out of its existing resources to catch up with the frontier. Statement: This is an exclusive Sina.com article. Reproduction without authorization is prohibited. tech_web_article_v2015_block_bottom Non-targeted 300*250 button 17/09 wenjing begin

Non-targeted 300*250 button end

Source

Sina Tech