Tencent Hy4 Explained: What It Is, How It Works, Limits and Risks

Last updated: 9 November 2026

On 4 November 2026, Tencent announced Hy4, the full version of its new flagship large language model and the successor to the Hy4 preview it open-sourced in late August. Two things make this release stand out. First, it is the first model in Tencent’s Hy family to add visual understanding and computer use, meaning it can look at images and operate both terminals and graphical apps. Second, Tencent made an unusual safety claim: it says offensive hacking material was deliberately removed from the training data, so the model learns to find and fix vulnerabilities rather than to break into systems.

That second point arrives at an interesting moment. Over the past few weeks we have seen frontier labs hold back their strongest cyber-capable models, and downloadable Chinese models score surprisingly close to restricted ones on offensive security tests. Tencent is betting that you can release an open-weight model responsibly by shaping what it learns in the first place. Then, on 7 November, Tencent’s own chief AI scientist warned publicly that AI biology risks are approaching a dangerous threshold, which added a second layer to the conversation.

In this post I explain what Hy4 actually is, what is confirmed and what is still only promised, how its Mixture-of-Experts design works, what the only independent benchmark so far says, the safety and privacy questions, and why I think Indian students preparing for GATE DA, CS and cybersecurity roles should follow this release closely, even if they never run the model themselves.

Quick summary

• What it is: Tencent’s flagship large language model, announced on 4 November 2026, with reasoning, coding, agent, vision and computer-use abilities.
• The preview (open now): Hy4 preview, released on 28 August 2026 under the Apache License 2.0, is a Mixture-of-Experts model with 770 billion total parameters, about 49 billion active per token and a context window of roughly 1 million tokens.
• The full model (partly open): Available through Tencent’s own apps and APIs and through OpenRouter. Tencent says the weights and system cards will be released “in the coming days”. When I checked on 9 November, I could not find a public repository for the full Hy4 weights, only the preview.
• Safety claim: Tencent says cyber training data was limited to defensive work (writing proofs of concept and patching vulnerabilities), with exploitation material excluded.
• Independent view: Artificial Analysis gives Hy4 a score of 35 on its Intelligence Index v5.0.1, placing it 40th of 69 models it lists, slightly below the median, and notes the model is very verbose.
• Price: Well under a dollar per million input tokens, but sources disagree on the exact figure (details below).

What is Hy4, in plain language?

Tencent is the Chinese technology company behind WeChat, QQ and a large gaming business. Its AI models were long branded “Hunyuan”; the current generation is branded simply “Hy”. Hy3 was the previous flagship. Hy4 is the next step up, and Tencent positions it as one of the leading open-source models in the world.

It helps to separate two products that share a name:

• Hy4 preview (August 2026): a text-only reasoning and coding model whose weights you can download today from Hugging Face, ModelScope, GitCode, CNB and GitHub, including an official FP8 (8-bit) build.
• Hy4 (November 2026): the finished model, which adds image understanding and computer use, ships with Tencent’s “Hyra 2.0” agent harness, and is already running inside Tencent products such as the Yuanbao assistant, WorkBuddy and CodeBuddy. Its weights are promised but, as of this writing, not yet public.

This distinction matters because many articles mix the two. The 770 billion parameter figure, the Apache 2.0 licence and the architecture details all come from the preview model card. Tencent’s November announcement does not restate a parameter count or a licence for the final model, and the independent tracker Artificial Analysis currently lists the final model’s size as not disclosed and its status as proprietary. It is reasonable to expect the final model to follow the preview’s design, but until Tencent publishes the weights and model card, treat that as an expectation rather than a fact.

How Hy4 works

Mixture of Experts

Hy4 preview is a Mixture-of-Experts (MoE) model. Instead of one giant feed-forward network that every token passes through, each MoE layer contains many smaller “expert” networks plus a small router that decides which experts each token should use. According to the official model card:

• The model has 78 layers. The first is an ordinary dense layer; the other 77 are MoE layers.
• Each MoE layer has 256 routed experts and 1 shared expert. For every token, the router picks the top 8 routed experts, and the shared expert always runs.
• That is why the model has 770 billion parameters in total but only about 49 billion active for any given token. You pay the memory cost of the full model, but the compute cost per token is closer to that of a 49B dense model.
• An extra multi-token prediction (MTP) layer of about 10 billion parameters lets the model draft several tokens ahead for faster “speculative decoding”. This is why Hugging Face sometimes shows the size as 780B.

Sparse attention for a 1 million token context

Standard attention compares every token with every earlier token, so its cost grows roughly with the square of the sequence length. A 1 million token context would be impractical that way. The preview card says Hy4 uses a gated variant of DeepSeek’s sparse attention technique, where a lightweight “indexer” first selects the most relevant 2,048 earlier tokens and full attention is computed only over those. It also compresses the keys and values it stores, which shrinks the memory needed for long conversations.

Reasoning by default

Hy4 is a reasoning model: by default it “thinks” before answering, and the preview card lets developers switch this off with a no-thinking setting for quick replies. Tencent’s own model card admits a known weakness here: the preview can spend longer than needed reasoning and tends to over-verify its own work. That matches what the independent testers found, as you will see below.

What changed in the full release

Tencent’s 4 November announcement lists the main upgrades: visual understanding for the first time in the Hy line; computer use across command-line and graphical interfaces (Tencent mentions training on professional software such as Blender and CAD tools); better agent behaviour inside its own products; and fewer hallucinations.

What Hy4 can do: the claimed numbers

Almost every performance figure for the final Hy4 comes from Tencent itself. The table below separates company claims from independent measurement.

A note on the hallucination numbers: Tencent’s English announcement and Interesting Engineering give a fall from 12.95% to 7.2%, while Digital Applied and several other reports give 15.76% to 5.45%. Both agree that “unsupported assumption” errors fell from 23.1% to 5.6%, so I have not put a single hallucination figure in the headline.

Some outlets also published a twelve-benchmark table for the final model (Terminal Bench, OSWorld and others). Those numbers were read off a chart image in Tencent’s post by third parties, and one of those write-ups itself notes that Hy4 did not rank first on any of the twelve. I have left them out of the table because I could not check them against a text source.

How to try Hy4 (official routes)

Use it without installing anything

Tencent says the full Hy4 is available through its TokenHub API platform, its WorkBuddy and CodeBuddy coding tools, the Yuanbao assistant and the ima knowledge app, and through third-party platforms including OpenRouter and OpenCode. Tencent’s announcement says WorkBuddy and CodeBuddy are free to use with Hy4 from 4 to 15 November, and that TokenHub offers 50% off until 31 December 2026.

On pricing, sources disagree slightly. Tencent’s English announcement says Hy4 keeps the same pricing as the preview at roughly US$0.90 per million input tokens and US$2.70 per million output tokens. Digital Applied and Artificial Analysis list US$0.834 and US$2.501 (about ¥6 and ¥18 in Tencent’s Chinese pricing), which are the figures Tencent gave for the preview. The difference is probably rounding or currency conversion. Either way, it is far cheaper per token than the top US closed models, but remember that a verbose reasoning model uses many tokens per answer.

Run the open preview yourself

Only the preview weights are downloadable today. The model card recommends vLLM or SGLang for serving and gives Docker-based commands. A trimmed version of the official vLLM example looks like this:

# Official vLLM example from the Hy4 preview model card (trimmed)
docker run --gpus all -p 8000:8000 --ipc=host \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:hy4-preview tencent/Hy4-preview-FP8 \
  --tensor-parallel-size 8 \
  --served-model-name hy4-preview --port 8000

Please read the full command on the official model card before running it, because it also sets speculative decoding, tool-call parsing and attention options.

Be realistic about hardware. Even in 8-bit form, 770 billion parameters need roughly 770 GB just to hold the weights, before any memory for long contexts. The official example splits the model across 8 GPUs, and those need to be very high-memory data-centre GPUs. This is not something a student can run on a laptop or a single cloud GPU; for learning, use the API or study the model card.

Limits, criticism and independent results

• Middle of the pack on independent testing. The Artificial Analysis score of 35 puts Hy4 below the median of the models it compares. That is a respectable open model, not a frontier leader.
• Very verbose and slow to start. Artificial Analysis measured about 160 million output tokens to complete its index, against a median of about 73 million, and a time to first answer token of about 103 seconds on Tencent’s API, against a median near 15 seconds. Its output speed of about 94 tokens per second is fine once it starts. Other users report faster first responses on OpenRouter, so latency depends on the provider.
• Mostly self-reported benchmarks. The blind expert reviews, the finance study, the WorkBuddy scores and the hallucination figures are all Tencent’s internal tests. TechNode noted when covering the preview that its results had not been independently verified, and that remains true for the full model.
• “Open source” is still a promise for the final model. Reuters reported that weights and a safety report would come “later”; Tencent’s own post says “in the coming days”. Digital Applied found the Hugging Face and GitHub addresses for the full model unavailable on 5 November. Until the weights appear with a licence, you cannot inspect, fine-tune or self-host the November model.

Safety: the defensive-only cyber training claim

This is the most interesting part of the release for anyone in security. Tencent says Hy4 was trained on defensive security data, focused on writing proofs of concept and fixing vulnerabilities, and that “offensive materials related to exploitation are explicitly excluded from the training dataset.” Reuters reported that Tencent excluded material on building hacking tools, breaking into systems and probing systems for weaknesses. Tencent’s Xuanwu Lab and Keen Security Lab red-teamed the model and, according to Tencent, found no significant increase in offensive risk.

My view: filtering training data is a sensible, under-used safety lever, and I would like to see more labs try it. But three caveats apply. First, defensive and offensive knowledge overlap heavily; you cannot write a good patch without understanding the bug, and a proof of concept is close to an exploit. Second, the red-team results are internal, and the full system card has not been published. Third, once open weights are out, anyone can fine-tune the model on the excluded material. Data filtering raises the bar; it does not lock the door.

On biology, Tencent says Hy4 has strong basic biological knowledge but is more limited than frontier models at open-ended experimental design, and reports a score of 30.4 out of 100 on its internal biology experiment benchmark, below its own high-risk threshold. Separately, on 7 November, Tencent’s chief AI scientist Yao Shunyu told a Hong Kong science event, as reported by the South China Morning Post and The Next Web, that AI capability in biology will soon reach a point where it is “actually possible to do harm”, and that Chinese open-model developers must take safety very seriously. It is notable to hear that from inside a company releasing open weights.

Privacy and the Indian angle

This is general information, not legal advice.

If you use Hy4 through Tencent’s hosted services, your prompts and files are processed by Tencent under its terms and the laws that apply to it. India’s Digital Personal Data Protection Act, 2023 places duties on organisations that process personal data, and many Indian universities, coaching institutes and companies have their own rules about which external AI services may receive student or customer data. Before pasting personal data, answer sheets or company code into any hosted AI service, Chinese or American, check those rules.

Open weights change this picture. A self-hosted copy of an open model running on Indian infrastructure keeps the data with you, which is one reason open-weight models matter to Indian institutions and startups. The catch with Hy4 is the hardware bill and the fact that the full model’s weights are not yet out. Also note that a Chinese model may reflect content restrictions that apply in China; test it on your own use case before relying on it for sensitive topics.

Why this matters for Indian students and educators

You do not need to run a 770 billion parameter model to learn from this release. Hy4 is a compact case study in several syllabus topics:

• Machine Learning (GATE DA): Mixture of Experts, gating, sparse activation and the trade-off between total and active parameters.
• Algorithms and Data Structures: why quadratic attention fails at 1 million tokens and how top-k selection makes it tractable.
• Computer Organisation and Architecture: memory bandwidth, 8-bit (FP8) number formats, and why model size is a hardware problem before it is a software one.
• Operating Systems and Distributed Systems: tensor parallelism across 8 GPUs, batching and serving with vLLM and SGLang.
• Probability and Statistics: reading benchmark claims critically: blind reviews with internal experts, sample sizes, and the gap between self-reported and independent results.
• Cybersecurity: the defensive versus offensive distinction, responsible disclosure, and how AI safety policy is becoming a technical discipline in its own right.

For careers, the skills that grow in value are the ones this release exposes: evaluating models honestly, serving large models efficiently, and building safety evaluations. Indian AI startups, GCCs and research labs are hiring for all three. A good exercise for a final-year project: pick one open model, write ten tasks from your own domain, and compare its answers and token usage with a smaller model. That is a small version of exactly what Artificial Analysis does.

Frequently asked questions

Is Tencent Hy4 open source?

The Hy4 preview is open-weight under the Apache License 2.0 and can be downloaded today. For the full November model, Tencent has promised weights and system cards but, when I checked on 9 November 2026, they were not yet publicly available.

How big is Hy4?

The preview has 770 billion total parameters with about 49 billion active per token, plus a roughly 10 billion parameter multi-token prediction layer. Tencent has not restated the size for the final model.

Can I run Hy4 on my laptop or a single GPU?

No. Even the 8-bit version needs hundreds of gigabytes of GPU memory; the official serving example uses 8 data-centre GPUs. Use the API or a hosted platform instead.

Is Hy4 better than GPT, Claude or Gemini models?

Not on independent testing so far. Artificial Analysis ranks it 40th of 69 on its intelligence index. Tencent’s own comparisons are mostly against other Chinese open models, where it is competitive.

Does removing offensive data make Hy4 safe for cybersecurity work?

It reduces the risk of easy misuse, according to Tencent’s internal tests, but it is not a guarantee. Defensive and offensive knowledge overlap, the system card is not yet published, and open weights can be fine-tuned by anyone.

Is my data safe if I use Hy4 through Tencent’s apps?

Your data goes to Tencent’s servers under its terms. Follow your institution’s AI policy, avoid sharing personal or confidential data with any hosted AI service, and prefer self-hosted open models when data protection matters.

The bottom line

Hy4 is a serious, cheap and capable open-weight contender from one of China’s biggest technology companies, and its defensive-only cyber training is a genuinely interesting safety experiment. It is not, on current independent evidence, a frontier leader, and the full model’s weights and safety report are still a promise rather than a download. My advice to students: study the preview’s architecture because it teaches modern LLM design beautifully, treat company benchmarks with healthy scepticism, and watch whether Tencent publishes the weights and system card as promised. How it handles that step will tell us more about responsible open AI than any benchmark.

Sources and further reading

• Tencent: Tencent launches Hy4 (4 November 2026)
• Tencent: Tencent releases and open-sources Hy4 preview (28 August 2026)
• Hugging Face: tencent/Hy4-preview model card
• Reuters (via BusinessDay): Tencent releases new AI model, pledges safeguards against cyberattack use
• Artificial Analysis: Tencent Hy4 model page
• Interesting Engineering: Tencent Hy4 challenges world’s top models
• Digital Applied: Tencent Hy4 is out, what is confirmed and what is not yet
• The Next Web: Tencent’s Yao Shunyu on AI biology risk
• TNGlobal: Tencent open-sources 770B Hy4 preview model

Written by Piyush Wairale (M.Tech, IIT Madras), founder of PiyushAI Edtech. Model availability, prices and benchmark scores change quickly; figures are as reported on or before 9 November 2026.