Tencent Releases Hy4 With Vision and Computer Use but Holds Back Model Weights

Tencent released Hy4 on Wednesday, November 4, 2026, calling it the full version of its Hunyuan flagship large language model, and for the first time the system can read images and operate software interfaces on its own. The Shenzhen company also said the downloadable model weights and a detailed system card would follow in the coming days rather than arrive with the launch, a sequencing choice that separates Hy4 from the open-weight framing Tencent has used for earlier releases.

Tencent Holdings is the Chinese technology group behind WeChat, the country's most popular app with more than 1 billion monthly active users, and its Hunyuan team has spent the past two years building models that it distributes both as hosted services and, at times, as downloadable files. The company has positioned its artificial intelligence work as an enterprise play, plugging models into workplace and coding tools that sit close to the businesses that pay for cloud computing.

The previous Hy4 Preview arrived in August 2026 as an open-sourced release, and Tencent described it as having 770 billion total parameters, 49 billion active parameters and a context window exceeding one million tokens. The full Hy4 keeps the same family name but changes the delivery model at launch: hosted APIs and Tencent products first, files later. That distinction matters in a market where Chinese labs including DeepSeek, Z.ai and Alibaba have competed loudly over open weights, and where the phrase open source has become a marketing claim as much as a licensing term.

The launch also lands in a heated safety debate. Regulators and security researchers have spent the past two years warning that capable models could lower the barrier to cyberattacks, and Tencent itself disclosed last month that AI agents it used to test third-party models displayed unintended web-browsing behavior, though it said the activity did not come from its Hunyuan models and found no evidence of deliberate targeting or coordination.

Key Facts

Reuters reported on November 4 that Tencent designed Hy4 with limits intended to reduce the risk that it could be used for cyberattacks or harmful biological research. The company restricted cybersecurity training to defensive uses such as identifying software flaws and helping fix them, and excluded material on building hacking tools, breaking into systems and testing systems for weaknesses. Tencent said internal security specialists tested Hy4 and found no sign that it posed higher risks in helping users break into software or bypass protections, and that testing found the model was not capable of meaningfully making harmful biological experiments easier to carry out.

The capability list is broader than the preview's. Tencent said Hy4 introduces visual understanding for the first time, supporting fine-grained image processing, complex chart analysis and mathematical visual reasoning, and introduces computer-use abilities that operate across command-line and graphical interfaces. Training covered professional software including Blender and computer-aided design tools. The model is available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's Dayuan, ima, OpenRouter, Hermes and OpenCode.

Tencent published a set of performance claims alongside the launch. On WorkBuddy, the company reports an average score of 83.18 and a pass rate of 96.15%, ranking first in nine of twelve scenarios in mainland China. In a blind financial evaluation with 213 finance professionals, including 153 external experts, about 70% of whom had more than five years of experience, across 579 real-world tasks, Hy4 scored 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3. Tencent says the hallucination rate fell from 12.95% in the preview to 7.2%, and errors from unsupported assumptions dropped from 23.1% to 5.6%.

Unite.AI reported on November 4 that listed API pricing is unchanged from the preview at $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached input tokens, while Tencent's own announcement gave approximately $0.90, $2.70 and $0.045 for the same three categories. OpenRouter's deployment listing specifies a 1,048,576-token context window and a 65,536-token completion limit, accepts text and image inputs and returns text, with Tencent Cloud as the only listed provider. Tencent said WorkBuddy and CodeBuddy would be free from November 4 to 15 and TokenHub would be 50% off through December 31, 2026.

Data Phoenix reported on November 5 that the launch was hosted-only, with weights, license and system cards still unpublished, and that access through Hermes and OpenCode was not separately verified, although an OpenRouter listing was live. The report noted that Tencent's November 4 announcement did not confirm the 770 billion and 49 billion specifications for the final model or state the release license, and that no firm publication date or repository address was given for the weights, the Flash variant or the system cards.

Analysis

The bigger picture here is that Tencent wants two reputations at the same time: the responsible lab that trains its model away from offensive cyber capability, and the aggressive challenger that can match DeepSeek, Z.ai and Alibaba on capability per dollar. Those goals pull in opposite directions. Restricting training data to defensive security work is a real constraint, not a press release flourish, because it removes exactly the material that makes a model useful for penetration testing, and Tencent accepts that trade in exchange for a cleaner safety story it can tell regulators and enterprise buyers.

The safety claims deserve scrutiny. Tencent reports a score of 30.4 out of 100 on Hy-BioExperimentBench, a number the company presents as evidence of limited biological capability, and says its internal specialists found no elevated risk of helping users break into software. Those are company claims. Data Phoenix reported on November 5 that the opened sources did not include an independent laboratory evaluation of the full model, which means the strongest public evidence for the safety posture is Tencent's own testing, conducted by the same organization that benefits from a favorable result.

Withholding the weights is the most interesting decision. Releasing the preview in August as an open-sourced model bought Tencent goodwill and a developer base; shipping the flagship hosted-only lets the company capture paying API traffic first, keep a license option open, and stage the safety documentation on its own schedule. What this really means is that open weights have become a product decision rather than an ideological commitment. Tencent can still call Hy4 an open-weight model in its marketing while the files do not exist.

The capability gap is real and Tencent admits it. The South China Morning Post reported on November 4 that Tencent acknowledged Hy4 still trailed the world's leading closed-source systems, particularly on challenging internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering, even as the company said the release marked the first time its lineup became one of the strongest open-weight models available today. That is a narrower claim than leadership, and it is the honest one.

Why It Matters

For enterprise buyers, Hy4 matters because of where it sits rather than what it scores. The model is wired into WorkBuddy, CodeBuddy, Yuanbao, WeCom's Dayuan and ima, the tools Chinese companies already use for documents, code and internal chat, and the promotional pricing makes the first fifteen days cheap. A model that can read a chart, operate Blender or CAD, and hold a million-token context is useful in exactly the workflows for which Tencent sells cloud capacity.

For the wider market, the launch sharpens a split between Chinese labs that publish weights and those that rent access. If Tencent's flagship stays hosted, the open-weight race with DeepSeek, Z.ai and Alibaba becomes a race among secondary releases, with the best systems kept behind APIs. That would push developers toward whichever vendor offers the best price for capability and make switching costs the deciding factor rather than licensing freedom.

For safety policy, the precedent is that a major lab can pre-announce open weights without a date, publish partial evaluations, and still be described as open. Regulators looking for disclosure standards will find little to verify until the system card appears. The norm being set is not secrecy, it is delay with good intentions attached.

Next Up

Tencent said Hy4 Flash is scheduled for release soon, and that the full weights and system cards will follow in the coming days, but it gave no firm publication date or repository address. vLLM and SGLang support are planned alongside the weights, according to Unite.AI. Until those files land, every benchmark figure above remains a vendor claim.

The next real test is whether the open release, when it arrives, carries the same 770 billion parameter and 49 billion active parameter specifications as the August preview, and what license Tencent attaches. The company did not confirm those numbers for the final model. Watch the system card as closely as the weights: it is where the cybersecurity and biological training limits will either be documented in detail or left as assertions.