What did Tencent release on 4 November?
Tencent released Hy4 on Wednesday 4 November 2026. The company called it a major step forward and its first model to rank among the strongest open-source systems. The release replaces Hy4 preview, which Tencent published on 28 August and kept in service at the same API price.
Hy4 keeps the preview's mixture-of-experts design. TechNode reports a total of 770 billion parameters. Preview documentation describes 49 billion active per token, 78 layers, 256 routed experts and a native multi-token-prediction layer for speculative decoding. OpenRouter lists a one-million-token context window for the hosted model.
Vision is the clearest addition. Tencent says Hy4 can read fine-grained image details, interpret complex charts and reason across multiple images. It also supports computer use for the first time. The model works through command-line and graphical interfaces and follows long workflows that span several applications.
Tencent trained the model on specialist software such as Blender and computer-aided design tools. It also adapted Hy4 for Unreal Engine, Unity and Godot, and extended it to multiplayer game development.
Hy4 is available through Tencent Cloud and OpenRouter, and inside CodeBuddy, WorkBuddy, Yuanbao and ima. The South China Morning Post reports that the launch also brings a broad update to WorkBuddy, Tencent's workplace assistant. The WorkBuddy integration with WeCom connects to chat history, email, documents, calendars and meeting notes.
Tencent did not raise prices. The listed API rates match the preview. They are $0.042 per million cached input tokens, $0.834 per million input tokens and $2.501 per million output tokens. TechNode reports that CodeBuddy and WorkBuddy users receive a limited free quota from 4 to 15 November.
The vision and computer-use features matter because they move Hy4 from a chat model into an agent. A model that can read a screenshot and click through an interface can take over repetitive software work. That is the agent use case behind much of the current interest in computer-use models. Whether Hy4 does that reliably is a separate question from whether it can do it at all.
Why does Hy4 matter in the open-weight race?
China's largest AI labs have spent the past year releasing open-weight models aimed at closing the gap with proprietary systems. The South China Morning Post frames Hy4 as part of a race involving DeepSeek, Z.ai and Alibaba. Tencent's own comparisons name Moonshot's Kimi K3.1 Pro and Z.ai's GLM-5.3 as peers.
Tencent is a newer entrant at that level. Nikkei Asia reports that the company has moved from proprietary models to open releases in order to win cloud customers. Hy4 is the clearest test of whether that strategy can produce a frontier-class model.
The economics explain the urgency. Open weights let developers test a model without a sales contract, and they let enterprises keep sensitive data inside their own infrastructure. Tencent still earns when those teams rent GPUs or buy hosted inference. That is why the company can give the model away and still expect a return.
Price is part of the pitch. At $0.834 per million input tokens, Hy4 costs the same as its preview. Its one-million-token context targets long documents, large codebases and multi-step agent work. Buyers get the capability gains without renegotiating their API terms.
The competitive field is crowded, though. Artificial Analysis currently names DeepSeek V4.1 Flash 1017 and GLM-5.3 as the highest-intelligence open models on its comparison page. Tencent's claim to the top tier therefore rests on benchmarks it ran itself, at least until outside evaluators publish results.
Open releases also put pressure on prices across the market. When a capable model is free to download or cheap to call, proprietary vendors must justify their premium. Reliability, support and features matter more than raw benchmark scores in that market. Hy4's launch pricing reinforces that trend in China and, through OpenRouter, beyond it.
There is a strategic message as well. A Chinese platform company is arguing that openness and safety can coexist. That argument lands at a time when US labs are tightening access and regulators are asking harder questions about model release. Hy4 is meant to show that Tencent can compete on capability, cost and caution at once.
Is Tencent Hy4 actually open source?
Not yet. Tencent's launch post says the company plans to "open-source Hy4's model weights". It adds that it will publish a "comprehensive system card" at the same time. Both are promises about a future release, not descriptions of the current one.
At the time of writing, the weights were not publicly available. The Hugging Face page linked from OpenRouter's Hy4 listing returned a 404 error. A ModelScope query for the Hy4 repository failed, and GitHub's Tencent-Hunyuan organisation hosted Hy4-preview but no Hy4 repository. The preview's weights, by contrast, resolve normally.
Several outlets still described Hy4 as open-source or open-weight on launch day. Reuters called it a new open-source model, and TechNode described it as a 770-billion-parameter model with weights due shortly. The label is directionally right but premature: downloadable weights are the thing that makes a model open.
Tencent says it is preparing the release with the vLLM and SGLang communities, and it names Inferact Inc. alongside vLLM. Those projects provide the serving software most teams use to run large models. Their support matters because a 770-billion-parameter model is not something a small team runs on one machine. The preview's deployment recipe asks for eight GPUs in parallel.
The preview shipped under Apache 2.0, according to its GitHub repository and Hugging Face metadata. Tencent has not said whether the final weights will use the same licence. That choice will decide whether startups can build commercial products on Hy4 without asking permission. It will also determine whether the release comes with usage limits.
A staged release is becoming normal in China. Reuters reported that Moonshot limited the cyber-offensive capabilities of its newest Kimi model in October. Chinese labs have joined US companies in trying to reduce loss-of-control risks. Tencent's decision to hold back the weights follows a similar pattern. The sequence is to publish the API, study how the model behaves, then decide how much to release.
Tencent's caution may also reflect how hard a release is to retract. Once weights are public, they can be copied, fine-tuned and redistributed. A system card published after the fact cannot undo a bad release. Companies therefore have an incentive to test first and publish second, even if that frustrates developers who want to self-host immediately.
How safe is Hy4 for cyber and biology work?
Tencent says it restricted what the model learned about offensive security. Training data was limited to defensive material on proof-of-concept generation and vulnerability patching. The company says it filtered out exploit development, reverse engineering and penetration testing.
Reuters summarised the same disclosure. The model was trained with limited cybersecurity data, restricted to "defensive" content, to reduce the risk that it helps with cyberattacks. The agency linked the caution to a wave of AI-driven cyberattacks that has raised concerns about advanced models and internet infrastructure.
Tencent also ran internal red-team audits. Researchers from Xuanwu Lab and Keen Security Lab tested cryptanalysis, reverse engineering and vulnerability exploitation. Tencent says their findings show no elevated risk in reversing cryptographic algorithms, defeating binary obfuscation or exploiting vulnerabilities.
For biological risks, Tencent reports a score of 30.4 out of 100 on Hy-BioExperimentBench. It describes that result as well below its high-risk threshold. The company says its experts reviewed the model's reasoning. They found no evidence that Hy4 meaningfully lowers the barrier to causing biological harm. It credits a gap between knowing biology and turning that knowledge into workable experimental protocols.
Those results come from the company that built the model. No independent audit has been published, and the system card that would explain the tests, prompts and thresholds is still pending. Readers should treat the safety numbers as claims to verify rather than settled findings.
Open weights would change the picture. Filters used during training stay with the weights, but safety systems added to a hosted API do not. Once the files are downloadable, anyone can run the model without Tencent's guardrails. That is why the timing of the release and the detail in the system card carry real weight.
How good is Hy4, and can we trust the benchmarks?
Most of the published numbers are Tencent's own. In a finance evaluation built with Tencent Financial Technology, 213 finance professionals rated outputs across 579 tasks. Hy4 scored 3.04, against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for both Kimi K3.1 Pro and GLM-5.3.
Tencent says 153 of the evaluators came from leading financial institutions. More than 70% had at least five years of industry experience. The design is more rigorous than a simple leaderboard, but the task set and scoring rubric were built in-house. The result measures Hy4 against Tencent's definition of good financial work.
The company reports large gains in reliability. On its internal tests, the hallucination rate fell from 15.76% in Hy4 preview to 5.45%. Speculative errors dropped from 23.1% to 5.6%. Those are striking improvements, but they are measured on the same internal suite that Tencent uses to guide training.
On game development, Tencent says Hy4 improves 19.1% on GameDev and 9.8% on GameLogic compared with the preview. The gains follow targeted training for Unreal Engine, Unity and Godot. They matter for Tencent's own game studios as much as for outside developers.
The method has continuity. For the preview, Tencent ran a blind side-by-side test in which 163 internal experts rated 203 engineering tasks. The preview edged GLM 5.3 and Kimi K3. The final model's finance study follows the same approach with outside professionals, though Tencent still controls the task design.
Tencent is unusually blunt about the gaps. Its post says a notable gap remains between Hy4 and top proprietary models. The listed weak spots are research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering. That candour sets expectations below the sweeping claims that often accompany a flagship launch.
Outside evidence is thin so far. Artificial Analysis runs standardised evaluations across models. It still ranks DeepSeek V4.1 Flash 1017 and GLM-5.3 as the strongest open models on its comparison page. Hy4's standing will become clearer when independent testers publish results, and when developers report how it performs on their own workloads.
What does Hy4 cost, and where can you use it?
Tencent lists Hy4 at $0.042 per million cached input tokens, $0.834 per million input tokens and $2.501 per million output tokens. The figures match the preview, so the capability gains arrive without a price increase.
TechNode reports a 50% discount on Tencent Cloud's TokenHub through 31 December. OpenRouter shows the discounted rate of $0.417 for input and $1.251 for output, with Tencent Cloud as the only listed provider. A single provider means less routing choice and no fallback if that endpoint struggles.
OpenRouter's telemetry offers an early look at production behaviour. The listing reported about 54 output tokens per second and 99.99% uptime over three days. It also showed a 40.95% structured-output error rate. That is a reminder that a new endpoint can be fast and reachable while still failing a specific kind of request.
The one-million-token context is the practical draw for teams working with long contracts, codebases or research sets. Context size alone does not guarantee good recall, and Tencent's long-context claims come from its own tests. Teams should measure retrieval and reasoning quality on their own documents before committing.
Access is broadest inside Tencent's products. CodeBuddy and WorkBuddy serve developers and office workers, Yuanbao reaches consumers and ima covers document work. Outside China, Tencent Cloud and OpenRouter are the main routes, and self-hosting is impossible until the weights are published.
For enterprises already using Tencent Cloud, Hy4 is a cheap upgrade with no migration. For teams standardised on AWS, Azure or Google Cloud, the model is another API to evaluate. The missing weights also remove the option of running it in their own environment. That limits its appeal to regulated sectors that cannot send data to a third party.
What should you watch next?
The first checkpoint is the weight release. Tencent has promised the model files and a system card but has not given a date. The licence it chooses will decide whether companies can deploy Hy4 commercially, and whether the open-weight label holds.
The second is Hy4 Flash. Tencent says a cheaper, faster variant is coming soon, with closer co-design for its products. If Flash keeps most of the capability at a lower price, it may matter more to high-volume workloads than the flagship.
Independent evaluation is the real test. Watch whether Artificial Analysis, LMArena and agent benchmarks add Hy4. Their scores will show whether they support Tencent's finance and gaming claims. A strong showing would validate the strategy; a weak one would suggest the internal benchmarks flatter the model.
Enterprise adoption will show up in Tencent Cloud usage and in how quickly free users convert to paid tiers after 15 November. Tencent has not disclosed revenue from Hy models, so the first public signals will come from product changes and cloud disclosures.
The larger question is whether gated releases become standard. Tencent trained a flagship on restricted cyber data, red-teamed it internally and held the weights back. If rivals copy that sequence, open-weight will increasingly mean downloadable after review. Launch-day benchmarks will then describe a model the public cannot yet inspect.