The new flagship ranks second among China-based labs, ties Grok 4.7, and still trails Moonshot AI's Kimi K3.1 Pro by a wide margin
Tencent’s newest flagship AI model, Hy4, scored 35 on the Artificial Analysis Intelligence Index v5. That makes it the second-highest-ranked model from any China-based AI lab.
Where Hy4 lands on the leaderboard #
The model that beats Hy4 among Chinese labs is Moonshot AI’s Kimi K3.1 Pro, which posted a score of 46 on the same index. That is an 11-point lead, a margin wide enough to fit a few other models inside it.
Hy4’s score of 35 puts it level with Grok 4.7 running at its xhigh setting.
The Artificial Analysis index rolls a range of evaluations into one composite number.
Tencent fully released Hy4 on November 4, 2026. A preview version arrived earlier, on August 28, 2026, giving the company roughly two months of real usage before the full launch.
Independent evaluations rank Hy4 well against peers from other prominent labs. It still sits behind the leading closed-source models built outside China.
The numbers behind the score #
On FrontierCode 2.0, a benchmark focused on agentic coding, Hy4 scored 31%.
Hy4 also took on a 579-task finance evaluation and scored 3.04. That result came in ahead of DeepSeek V4.1 Flash along with other competitors in the test.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Then there is pricing. Hy4 costs approximately $0.834 per million input tokens and $2.501 per million output tokens.
That pricing structure gives Hy4 a favorable cost-to-capability ratio, according to the research findings.
During its evaluation, Hy4 generated 160 million output tokens, which points to a talkative model.
Verbosity matters because output tokens are billed. A cheap per-token rate can look less cheap if the model insists on writing the long version of every answer.
What changed since the preview #
The full release adds several capabilities the August preview lacked. Hy4 now handles vision, meaning it can work with images alongside text.
It also supports tool calling and computer use for the first time. Tool calling lets a model trigger external functions, such as querying a database. Computer use lets it operate software interfaces more directly, closer to how a person clicks through a screen.
The context window grew to 1,048,576 tokens, or about one million.
Hy4 is described as Tencent’s strongest open-weight model to date. Tencent is also preparing to release model weights and system cards, which would give outside researchers a closer look at how the model was built and how it behaves.
Tencent has leaned on real-world feedback from its broad product ecosystem to shape the model. That loop shows up most clearly in Hy4’s results on long-horizon agentic work and practical applications.
Analysts from Goldman Sachs pointed to this closed-loop approach as a key differentiator for Tencent.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.