Tencent's Hy4 scores 35 on independent AI index, tying Grok 4.7

One day after Tencent released the full Hy4, the model has its first independent composite score: 35 on the Artificial Analysis Intelligence Index, level with Grok 4.7 and second among models from China-based labs. At the same time, Tencent is giving more detail on a safety choice made during training. It says Hy4’s cybersecurity training is limited to defensive uses, and offensive hacking material is excluded.

This matters because our launch report found that nearly all of Hy4’s performance figures came from Tencent’s own internal tests. The Artificial Analysis number is the first broad third-party measure of the release. The safety claims, by contrast, still rest on Tencent’s own word.

Where Hy4 lands on the Artificial Analysis index

Artificial Analysis, an independent evaluator, lists Tencent Hy4 at 35 on its Intelligence Index. It describes that score as above average among comparable models, where the median is 30, and ranks Hy4 43rd of 106 on the model’s page. The score is for the reasoning version of Hy4. Artificial Analysis notes that a non-reasoning variant may also exist.

The index is a composite: it combines many separate tests into one number. Version 5.0.1 combines nine evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset) and Terminal-Bench-Science 0.1. Many of these test agents, meaning models that take multi-step actions with tools. For example, Artificial Analysis describes FrontierCode 2.0 as agentic software engineering in open-source repositories, and Terminal-Bench 4.0 as agentic coding and terminal use.

Crypto Briefing reports these placements against rivals:

• Hy4’s 35 ties Grok 4.7 running at its xhigh setting.
• Among China-based labs, only Moonshot AI’s Kimi K3.1 Pro scores higher, at 46. That is an 11-point gap.
• Hy4 still trails the leading closed-source models built outside China.
• On FrontierCode 2.0, the agentic coding component, Hy4 scores 31%.

For context on the leader among Chinese labs, see our report on Moonshot’s Kimi K3.1 Pro weights release.

Metric (Artificial Analysis)	Hy4
Intelligence Index v5	35 (median 30)
Rank on model page	43 of 106
Output tokens used across the index	160M (median 75M)
Verbosity rank	94 of 104
Input / output	Text and image in, text out
Context window	1M tokens
Public API listed	None

A talkative model: why the token count matters

Artificial Analysis flags Hy4 as very verbose. It generated 160 million output tokens while running the Intelligence Index, against a median of 75 million. Tokens are the chunks of text a model reads and writes, and API providers usually bill per token. As Crypto Briefing puts it, a cheap per-token rate can look less cheap if the model writes the long version of every answer.

Builders should treat cost figures with care for now. Crypto Briefing quotes per-million-token prices for Hy4 and calls its cost-to-capability ratio favourable. However, Artificial Analysis’ own page lists the cost per index task as not available and shows no publicly available API. Until Hy4 is served by providers that Artificial Analysis can measure, its speed and cost per task remain unknown on that page.

Artificial Analysis also labels Hy4 a proprietary model for now, even though Tencent describes it as an open-weight release. The likely reason is timing. Tencent has said the weights and a system card are coming “in the coming days” and have not shipped yet.

What Tencent says it kept out of training

The second development concerns safety. According to Reuters, as published by Rappler, Tencent says it restricted Hy4’s cybersecurity training to defensive uses, such as identifying software flaws and helping fix them. In that account, Tencent excluded material on building hacking tools, breaking into systems and testing systems for weaknesses. Tencent’s press release, as summarised by PYMNTS, says the model is trained on defensive security data but not on offensive materials related to exploitation.

Tencent makes these further claims, none of them independently checked yet:

• Internal security specialists tested Hy4 and found no sign that it posed higher risks in helping users break into software or bypass security protections.
• Its testing found the model was not capable of meaningfully making harmful biological experiments easier to carry out.
• Hy4’s capabilities in experimental design are limited compared with frontier models.

“We recognize the potential safety risks posed by advanced cyber and biological capabilities, and we are committed to a more responsible approach to model training, open-sourcing, and release,” Tencent said.

Reports differ on what “defensive” includes

The accounts do not fully match. The Reuters report says Tencent excluded material on testing systems for weaknesses. That kind of authorised testing is usually called penetration testing. However, remio reads Tencent’s English announcement differently. It says the announcement describes analytical reverse engineering and penetration testing as included, with only offensive exploitation material excluded. Reverse engineering means examining how software works. remio cautions against reading the release as a claim that Hy4 received no cybersecurity knowledge. It also notes that a training policy describes what a developer meant to include, not what the finished model can actually do.

Why Tencent is stressing safety now

The timing has context. Reuters ties the release to growing global concern about AI-powered hacking. It also notes that last month Tencent disclosed that AI agents it used to test third-party models showed unintended web-browsing behaviour. Researchers said agents linked to Tencent Cloud had sought information from Alibaba’s Amap service. Tencent said that activity did not come from its Hunyuan models, that it found no evidence of deliberate targeting, and that it would tighten controls over its agents’ internet access.

Tencent is also not alone. Reuters reports that Moonshot AI limited its newest model’s cyber-offensive capabilities last month. remio concludes that cybersecurity training restrictions have become a release consideration across competing developers. Separately, PYMNTS reports that Oracle ended a $7 billion agreement to provide AI computing capacity to Tencent after US officials raised concerns about the deal.

What this means for builders

The Artificial Analysis score gives Hy4 a credible independent anchor. It is above the median and level with a known US model, though clearly behind Kimi K3.1 Pro among open-weight contenders from China. Tencent’s own figures remain the main evidence for its specialist claims. For example, Tencent’s 579-task finance evaluation put Hy4 at 3.04, ahead of DeepSeek V4.1 Flash-1017 at 2.97 and Kimi K3.1 Pro at 2.68. Tencent’s internal tests also report the hallucination rate falling from 12.95% in the preview to 7.2%. These are company numbers, not third-party results.

Several practical points follow:

• Budget for verbosity. Twice the median output-token count means per-task cost and latency deserve their own tests, separate from list prices.
• Test security workflows directly. remio argues that security teams need to know which requests get useful answers, which get limited help, and which are refused. A model that refuses too much pushes defensive work elsewhere. A model that refuses too little raises obvious risks.
• Separate hosted access from weights. A promise to publish weights is a release plan. It does not mean researchers already have the files.

For background on Hy4’s architecture and licence terms, see our August report on the Hy4 preview.

What to watch next

• Weights and system card. Tencent says these are coming shortly. Reuters reports that a safety testing report will be released with the weights. That report should show how Tencent separated useful analysis from offensive material.
• Independent red-teaming. Red-teaming means deliberately probing a model for harmful behaviour. Once the weights are out, outside researchers can test whether the cyber boundary holds up against realistic requests, which remio calls the next task for evaluators.
• Speed and cost data. Artificial Analysis will need a public API endpoint before it can fill in Hy4’s time and cost per task.
• Hy4 Flash. Tencent says a cheaper Hy4 Flash model is coming soon.