# Tencent Hy4: 770B Parameters, Apache 2.0 License [2026] > Source: > Published: 2026-11-04 14:41:18+00:00 Tencent announced **Tencent Hy4** on November 4, 2026, and the company says the model sits among the world’s leading open-source AI systems. According to the [South China Morning Post](https://www.scmp.com/tech/big-tech/article/3373227/tencent-unveils-hy4-open-weight-ai-race-deepseek-zai-alibaba-heats), the release lands as the open-weight race between Tencent, DeepSeek, Z.ai and Alibaba gets more crowded. The Hy4 preview is a 770 billion parameter Mixture-of-Experts model that activates 49 billion parameters per token and ships under the Apache License 2.0. That combination matters to engineers for a plain reason. A permissive license plus downloadable weights means a team can run, fine-tune and ship the model without asking Tencent for permission. Few models at this scale come with terms that clean. This analysis separates what Tencent has confirmed from what circulates in headlines. It covers the architecture, the access channels, the related Hy4 Flash model and the competitive picture. Readers tracking the broader field can compare it with our earlier look at [Alibaba’s Qwen4 plans and the doubts around them](https://shattered.io/qwen4-launch-alibaba-open-source-promise-doubt-470-models-2026/), and with the wider [AI and machine learning coverage](https://shattered.io/ai-machine-learning/) on this site. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## What Tencent Announced on November 4 Tencent framed the news as the arrival of Hy4 among the leading open-source AI models. The official wording describes a [large language model](https://shattered.io/category/ai-machine-learning/) built on a Mixture-of-Experts design, with visual understanding as one of its headline abilities. The [Tencent newsroom](https://www.tencent.com/newsroom/all-news/?LanguageId=1) is the place to check for updates as the rollout continues. The naming deserves a quick explanation. The model Tencent calls Hy4 has an officially named preview, Hy4 preview, which carries the weights developers can download today. A related model, Hy4 Flash, targets buyers who care more about cost than peak capability. Tencent describes Hy4 Flash as carrying over Hy4’s visual-understanding and agentic capabilities. One caveat shapes the whole story. The phrase about joining the ranks of the leading open-source models is a characterization from Tencent and from the headline. The retrieved sources do not establish a precise ranking, and they publish no benchmark table that places Hy4 at a specific position. Readers should treat the ranking claim as Tencent’s own positioning until independent evaluations arrive. ## Hy4 Preview Specs at a Glance The figures below come from Tencent’s announcement and from the model’s GitHub repository. They describe the preview release, not a later general-release model whose parameter count Tencent has not confirmed. | Spec | Hy4 preview | |---|---| | Developer | Tencent | | Model type | Large language model, Mixture-of-Experts | | Total parameters | 770 billion | | Active parameters per token | 49 billion | | Context window | More than 1 million tokens | | Layers | 78 (1 dense feed-forward, 77 MoE) | | Experts per MoE layer | 256 routed plus 1 shared | | Experts active per token | 8 routed plus the shared expert | | License | Apache License 2.0 | | Weights released | Hy4 preview and Hy4 preview-FP8 | The [Hy4-preview repository on GitHub](https://github.com/Tencent-Hunyuan/Hy4-preview) carries the architecture details. It lists the repository date as August 27, 2026, while Tencent’s preview release is dated August 28, 2026. That one-day gap likely reflects time zones or upload timing, and nothing in the sources suggests two separate releases. ## Inside the 770B Mixture-of-Experts Design A Mixture-of-Experts model splits its feed-forward layers into many small expert networks. A router picks a handful of experts for every token, so most of the network sits idle on any given step. Hy4 follows that pattern with 256 routed experts and one shared expert in each MoE layer, and each token activates eight routed experts plus the shared one. The arithmetic is simple. Forty-nine billion active parameters out of 770 billion total means roughly 6.4% of the weights do work on a given token. Across the 77 MoE layers, the model holds 19,712 routed experts in total. Those are our calculations from Tencent’s published numbers, not figures Tencent itself stated. ### Why the Dense First Layer Matters The first of the 78 layers uses a dense feed-forward network, and the remaining 77 use MoE layers. Many MoE designs keep an early dense layer because routing decisions made on raw embeddings tend to be noisy. Tencent has not explained its reasoning in the sources we retrieved, so this remains an informed reading of a common design habit rather than a confirmed motive. The shared expert plays a different role. Because every token passes through it, it can hold knowledge that applies everywhere, which leaves the routed experts free to specialize. This arrangement is well established in the open-weight field, and Hy4 adopts it rather than inventing it. ## The Million-Token Context Window Tencent says Hy4 handles more than 1 million tokens of context. That puts it in the territory where a model can take in a large codebase, a stack of contracts or a long-running agent transcript in one pass. For coding assistants and document workflows, that range removes a lot of chunking and retrieval plumbing. A long window is a spec, not a guarantee of quality. Models often degrade on retrieval accuracy deep inside very long inputs, and the retrieved sources contain no long-context benchmark for Hy4. Teams planning to feed it entire repositories should run their own needle-in-a-haystack and multi-document tests before committing. Memory cost is the practical limit. Serving a 770 billion parameter model with a million-token cache takes serious hardware, even when only 49 billion parameters fire per token. The FP8 weights Tencent published help here, since 8-bit weights cut the memory footprint compared with higher-precision formats. ## Visual Understanding and Agentic Skills Tencent’s announcement describes Hy4’s visual understanding in three parts: image processing, complex-chart analysis and visual mathematical reasoning. That list targets work like reading a financial chart, solving a geometry problem from a photo or parsing a dashboard screenshot. These are the tasks where text-only models fail outright. Hy4 Flash inherits the same visual and agentic abilities at a lower price point, according to Tencent. That detail signals a product strategy. Tencent wants one capability set available at two cost tiers, so a developer can prototype on the larger model and move volume traffic to the cheaper one. Some outlets and developer blogs go further and describe computer-use features and agent benchmark wins. Those claims did not appear in the official facts we could verify, so we leave them out. If Tencent publishes a system card with agent evaluations, that document will settle the question. ## Where Developers Can Use Hy4 Today Tencent lists a broad set of entry points. Enterprise buyers can reach the model through Tencent Cloud TokenHub, while consumers and office users can find it inside Yuanbao, ima and WeCom’s AI Assistant, Dayuan. Developer tools WorkBuddy and CodeBuddy also carry it. For API access outside Tencent’s own stack, the announcement names OpenRouter, Hermes and OpenCode. - **Tencent Cloud TokenHub:** cloud API platform, Hy4 available. - **WorkBuddy and CodeBuddy:** work and coding assistants, Hy4 preview free for two weeks after launch. - **Yuanbao, ima and WeCom AI Assistant (Dayuan):** Tencent products with Hy4 available. - **OpenRouter, Hermes and OpenCode:** third-party API access. - **Hugging Face, ModelScope, GitCode and CNB:** hosts for the Hy4 preview and preview-FP8 weights. Tencent also said the preview could be reached globally through WorkBuddy and CodeBuddy. The free window runs two weeks after launch. After that, pricing details matter, and Tencent had not published confirmed prices in the sources we retrieved, so we do not quote any. Some third-party posts cite per-token rates, but we could not tie them to an official page. ## Apache 2.0 and What It Lets You Do The Apache License 2.0 is one of the most permissive licenses in common use. It allows commercial use, modification and redistribution, and it includes an explicit patent grant. Some open-weight models ship under custom community licenses that cap monthly users or bar certain uses, so Apache 2.0 stands out as the friendlier option. For a startup, that removes a legal review step. For an enterprise, it removes a negotiation with a vendor. Teams can fine-tune the weights on private data, host them in their own region and keep the results, which is the main reason regulated industries prefer open weights over closed APIs. The license covers the code and weights Tencent published. It says nothing about the training data, which Tencent has not detailed in the sources we saw. Open weights are therefore not the same as a fully open training pipeline, a distinction researchers raise whenever a lab uses the open-source label. ## The Open-Weight Race: Tencent, DeepSeek, Z.ai, Alibaba The South China Morning Post’s headline names the rivals directly, framing Hy4 as another move in a contest with DeepSeek, Z.ai and Alibaba. Chinese labs have pushed hard on open weights through 2026, and each release raises the bar for the next one. Our own tracking shows the pace. DeepSeek [open-sourced a super-agent on October 31](https://shattered.io/deepseek-open-sources-super-agent-oct-31-2026/), and our earlier explainer on [open-weight models side by side](https://shattered.io/open-weight-ai-models-side-by-side-the-spec-sheet/) lays out how the specs differ. We cannot offer a verified head-to-head table with Hy4 against those models, because the retrieved sources contain no shared benchmark. Any table that did so would rest on guesses. What the sources do support is a structural comparison, shown below. | Question | Hy4 preview | What is verified | |---|---|---| | Open weights? | Yes, preview and FP8 builds | Confirmed by the GitHub repository | | Permissive license? | Apache License 2.0 | Confirmed by the GitHub repository | | Vision input? | Yes | Confirmed by Tencent’s announcement | | Cheaper sibling? | Hy4 Flash | Announced, details limited | | Published benchmark rank? | Not in retrieved sources | Unconfirmed | | Official price list? | Not in retrieved sources | Unconfirmed | ### How Hy4 Compares With Other Open Models The practical contrast is with rivals that already run in production. Fintech firm Revolut, for one, [cut AI coding costs 10x with a self-hosted GLM-5.1](https://shattered.io/revolut-self-hosted-glm-ai-coding-costs-10x-2026/), a case that shows what open weights deliver once a team commits to hosting. Hy4 has no comparable public deployment story yet. It needs one before buyers treat it as a default choice. ## Why the Cybersecurity Angle Is in the Headlines A report carried by [MarketScreener Hong Kong](https://hk.marketscreener.com/news/tencent-rolls-out-hy4-ai-model-trained-with-limited-cybersecurity-data-ce785fdfd88bf021) carries a headline saying Tencent rolled out Hy4 trained with limited cybersecurity data. The Economic Times ran a similar headline that adds hacking fears to the framing. We did not retrieve the body text of either piece, so we make no claim about what Tencent did or why. The framing fits a wider worry. Strong open-weight models can help attackers as well as defenders, and labs increasingly weigh that tradeoff before release. Whether Tencent restricted specific data is a question for the system card, not for headline inference. Security teams adopting Hy4 should still treat it like any new model. Test prompt-injection resistance, review how agent tools get permissions and log all tool calls. The license grants freedom to run the model, and it also hands the operator responsibility for safe deployment. ## Market Impact for Cloud and API Buyers Open weights at this scale pressure closed API vendors on price. When a 770 billion parameter model is free to download, hosting providers can undercut proprietary rates, and OpenRouter’s listing makes the model easy to swap into existing apps. Our breakdown of [OpenRouter’s 5.5% fee against Together and Fireworks](https://shattered.io/openrouter-pricing-vs-together-ai-vs-fireworks-ai-5-5-percent-fee-2026/) shows where those marketplace costs sit. For Tencent, the incentive is clearer on the cloud side. Releasing weights drives developers toward Tencent Cloud TokenHub, where the company can sell managed inference, and toward WorkBuddy and CodeBuddy, where it can sell subscriptions. Open weights serve as a funnel, and the free two-week preview feeds the top of it. The money behind these releases is enormous. Our look at [AI capex topping $740 billion](https://shattered.io/ai-capex-boom-how-long-can-it-last-740-billion-markets-2026/) explains why labs keep shipping models even as markets question the payoff. Open releases help spread the cost of that bet across a developer community. ## Historical Context: How Tencent Got Here Tencent’s model work has long run under the Hunyuan name, and the Hy4 preview repository still lives under the Tencent-Hunyuan organization on GitHub. The Hy4 name marks a new generation in the company’s flagship line, and Tencent describes adding visual understanding to that line as part of the shift. The preview-first sequence stands out for readers who follow release patterns. Tencent shipped the preview on August 28, 2026, then made the broader Hy4 announcement on November 4, roughly 68 days later. That gap gave outside developers time to test the weights, which is a sound way to catch problems before a headline launch. Compare that with how closed labs launch. They typically announce and ship on the same day, with no public weights to examine. Tencent’s approach lets the community audit the model, though it also means early criticism may already exist by the time the formal launch arrives. ## Hardware Math for Self-Hosting Hy4 Self-hosting is where the 770 billion parameter figure bites. Weight storage scales with total parameters, not active ones, so a MoE model needs memory for all 770 billion even though 49 billion do the work on each token. The table below shows our back-of-the-envelope estimates for weights alone, calculated from the published parameter count and standard bytes-per-parameter sizes. | Precision | Bytes per parameter | Approx. weight size | Basis | |---|---|---|---| | BF16 or FP16 | 2 | About 1.54 TB | Our estimate | | FP8 (Hy4 preview-FP8) | 1 | About 770 GB | Our estimate | | 4-bit quantized | 0.5 | About 385 GB | Our estimate, no official build | These numbers exclude the key-value cache, activations and framework overhead, all of which add real gigabytes. A million-token context inflates the cache sharply, so a deployment sized only for the weights would run out of memory on long inputs. Plan for headroom well beyond the raw checkpoint size. The sparse design helps on speed even when it cannot help on storage. Because only 49 billion parameters compute per token, throughput per GPU resembles a much smaller dense model once the weights sit in memory. Teams with a multi-GPU node and a good inference engine can serve many users at once, which is the economic case for running MoE models at all. Smaller teams have a cheaper route. Hosted APIs through Tencent Cloud TokenHub and OpenRouter skip the hardware bill entirely, and the free two-week preview in WorkBuddy and CodeBuddy covers early experiments. The decision to self-host should wait until usage volume justifies the cluster, and until independent tests confirm the model beats cheaper options on your workload. ## Release Timeline: From Preview to Launch The dates below come from the fact record Tencent and its repository provide. The sequence shows a slow, preview-led rollout rather than a single launch day. - **August 27, 2026:** date shown on the Hy4-preview GitHub repository. - **August 28, 2026:** Tencent’s Hy4 preview release date. - **November 4, 2026:** Tencent announces Hy4 and Hy4 Flash. - **About two weeks after launch:** the free preview window on WorkBuddy and CodeBuddy ends. Developers who grabbed the preview weights in late August have had about ten weeks of hands-on time before the formal announcement. Their feedback, posted on the repository and in community forums, will shape how the launch is received. Early adopters often find quirks that a polished announcement never mentions, from tokenizer oddities to tool-calling edge cases. ## Five Predictions for Hy4 These are our forecasts, not reported facts. We build them from the confirmed specs and from how comparable releases have played out this year. - **Independent benchmarks will land within weeks.** Open weights invite third-party testing, and results on coding, vision and long-context tasks will decide whether the leading-model claim holds. - **Hy4 Flash will draw more volume than Hy4.** Cheap, fast models win most production traffic, and Flash carries the same visual and agentic capabilities. - **Quantized community builds will appear fast.** The FP8 release gives hobbyists and small labs a starting point for smaller formats that fit on fewer GPUs. - **Tencent will extend free access after the two-week preview.** Free windows tend to roll into discounted tiers when usage looks healthy, though Tencent has announced nothing of the kind. - **Rival labs will answer with their own releases.** The pace documented in the Qwen4 and DeepSeek coverage suggests another major open-weight launch before the year ends. ## What Developers Should Do This Week Start with the free preview in CodeBuddy or WorkBuddy, since it costs nothing for two weeks. Run your own tasks rather than leaning on public leaderboards. A coding assistant that tops a benchmark may still stumble on your framework or your repository layout. Next, test through OpenRouter if you already route traffic there, because swapping a model name takes minutes. Compare cost, latency and output quality against your current model on 50 to 100 real prompts. Keep a log, since the results will matter when Tencent publishes official prices. Finally, if you plan to self-host, pull the FP8 weights from Hugging Face or ModelScope and size your hardware before committing. A 770 billion parameter checkpoint demands multi-GPU nodes. Teams without that budget should wait for Hy4 Flash details or for community quantizations. Our guide to the [best open-source LLMs of October 2026](https://shattered.io/best-open-source-llm-october-2026-kimi-glm-deepseek/) offers alternatives that may fit smaller setups. ## What Remains Unconfirmed Several details are missing from the official sources we retrieved. Tencent has not confirmed a specific price list, a benchmark score, a precise ranking, a training-compute figure or the parameter count of the general-release model. No named individual spokesperson appears in the material either. Some third-party posts describe computer-use abilities, agent test wins and per-token rates. Others say weights were not yet downloadable at launch. These accounts conflict with the GitHub repository, which shows the preview weights on public hosts, so we rely on the repository and Tencent’s own statements. The takeaway: Hy4 is a real, downloadable, permissively licensed 770 billion parameter model, and the claims beyond that need independent proof. Watch for the system card, third-party evaluations and the Hy4 Flash specifications. ## FAQ: Tencent Hy4 **What is Tencent Hy4?** It is a large language model from Tencent built on a Mixture-of-Experts design, with 770 billion total parameters and 49 billion active per token in the preview release. Tencent announced it on November 4, 2026. **Is Hy4 really open source?** Tencent released the Hy4 preview and Hy4 preview-FP8 weights under the Apache License 2.0 on Hugging Face, ModelScope, GitCode and CNB. The training data is not described in the sources we reviewed, so open weights is the more precise term. **How large is the context window?** Tencent says it exceeds 1 million tokens. We found no independent long-context test results yet. **Where can I try it?** Tencent lists Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom’s AI Assistant (Dayuan) and ima. API access runs through OpenRouter, Hermes and OpenCode. The preview is free on WorkBuddy and CodeBuddy for two weeks after launch. **What is Hy4 Flash?** Tencent describes it as a cost-effective model that carries over Hy4’s visual-understanding and agentic capabilities. We have no confirmed specs or price for it. **Does Hy4 beat DeepSeek, Qwen or GLM?** No verified benchmark in our sources answers that. Tencent says Hy4 ranks among the leading open-source models, and independent testing will test that claim. **How much does Hy4 cost?** Tencent has not published a price we could confirm. Some third-party posts quote rates, but we left them out until an official page backs them.