Tencent Hy4: Why the 770B Open Model Matters

Key takeaways

• Tencent Hy4 preview is a 770-billion-parameter mixture-of-experts model that activates about 49 billion parameters for each token and offers a context window above one million tokens.
• Tencent published model weights under Apache 2.0, making Hy4 unusually permissive for a model of its size, but “open” does not make it cheap or simple to run.
• The model targets coding, office work and scientific research; most launch benchmarks are still vendor-reported and need independent replication.
• Hy4 matters because it packages open weights, API access and Tencent product distribution into one launch, giving developers several routes to adopt the same model.

Tencent Hy4 preview is now available as an open-weight AI model with 770 billion total parameters, roughly 49 billion activated per token and a context window exceeding one million tokens. Tencent released the model on August 28, 2026, placed the code and weights in public repositories under Apache 2.0, and made it accessible through its own products and third-party APIs.

Everyone else is reporting the enormous 770B number; we are explaining the practical trade-off hidden inside it. Mixture-of-experts routing limits the computation used for each token, but a self-host still needs access to the full model weights and datacentre-class memory. Hy4 is open to inspect and deploy, yet most businesses will encounter it first as a rented API rather than a local download.

What exactly did Tencent release?

Tencent calls Hy4 preview a next-generation large language model built for coding, office productivity and scientific research. The company says the model contains 770 billion total parameters, with 49 billion active for each token, and can process a context longer than one million tokens. The “preview” label matters: Tencent is seeking real-world feedback before a later official release.

The weights are available from Tencent’s official repositories, including GitHub and Hugging Face, and the licence file states Apache License 2.0. Users can also reach the model through Tencent products such as WorkBuddy, CodeBuddy, Yuanbao and ima, or through Tencent Cloud TokenHub and OpenRouter. This multi-route distribution reduces the delay between a research release and actual product use.

Hy4 is a text model rather than a native image or video model. Tencent’s launch description focuses on software engineering, document-heavy office tasks and research. Developers should verify the specific endpoint, context allowance, price and data-handling terms of whichever hosting provider they choose because those operational details can differ from the model’s maximum configuration.

Tencent Hy4 is an open-weight 770B mixture-of-experts model, but it does not perform 770B parameters of computation on every token. Its router selects a smaller expert path of about 49B active parameters while the deployment still carries the full model.

Why the 770B and 49B numbers both matter

A model parameter is a learned numerical value used to transform input into output. Larger parameter counts can increase capacity, but they do not automatically prove better reasoning or reliability. Architecture, training data, post-training, inference settings and tools all influence results.

Hy4 uses a mixture-of-experts, or MoE, architecture. Instead of applying every model component to every token, a router selects a subset of specialised expert networks. Tencent’s reported 49B active figure describes the approximate parameters used in that path. This can provide the capacity of a very large model without the compute cost of running all 770B parameters for each step.

The saving has an important limit: active parameters are not the same as resident parameters. A self-host generally must store the entire set of weights and make them available across accelerators. Official deployment examples point to multi-GPU servers, not an ordinary laptop. Quantisation can reduce storage and memory requirements, but it adds another engineering decision and may affect quality or speed.

Hy4 figure	What it means	What it does not mean
770B total parameters	The capacity of the full model backbone	Every parameter runs for every token
49B active parameters	The approximate expert path used per token	Only 49B weights need to be stored
1M+ context	The configured maximum input-and-output window	Perfect recall across every long document
Apache 2.0	Permissive rights to use, modify and distribute	Zero infrastructure or compliance cost

Is Tencent Hy4 really open source?

Tencent uses the phrase “open-sources,” and its repository provides downloadable weights, code and an Apache 2.0 licence. Under common industry usage, that is a notably permissive open-weight release. The licence allows commercial use and modification with the required notices.

There is still a useful distinction between open weights and complete training transparency. Downloadable weights let a developer inspect, run and fine-tune the trained model. They do not necessarily include every training example, data-filtering decision or full recipe needed to reproduce the model from scratch. Businesses should describe the release precisely rather than assuming that “open” answers every provenance question.

The permissive licence can nevertheless change adoption. A company can build a private deployment, adapt the model for a domain or use a hosted version without being locked to one proprietary interface. For researchers, published weights make independent evaluation, safety testing and compression possible.

What Tencent claims about Hy4 performance

Tencent reports strong results across coding, office tasks and scientific research. Its launch materials compare Hy4 with open and closed rivals on software engineering, long-context understanding, tool use and reasoning. The company also says more than 160 experts participated in blind testing and preferred Hy4 over Hy3 in a large majority of evaluated scenarios.

Those are vendor-reported results. They are useful as a description of Tencent’s target, but they should not be treated as independent proof that Hy4 is the best model for a business. Benchmark prompts can be sensitive to tool access, reasoning effort, sampling settings and grading methods. A result produced with browsing or code execution is not directly comparable with a text-only run.

Independent evaluations published after launch show a mixed picture rather than a universal winner. Some analyses find large improvements over Hy3 and competitive coding performance, while also noting that closed frontier models lead many difficult reasoning rows. The responsible conclusion is that Hy4 appears competitive enough to test, not that a single launch chart settles the market.

What does a one-million-token context enable?

A context window is the amount of text and generated output a model can consider in one request. A window above one million tokens can fit large codebases, long reports or many documents. That makes it attractive for repository analysis, contract review, literature synthesis and agents that carry a long task history.

Capacity is not the same as accuracy. Long-context models can miss a detail, over-weight recent passages or make a confident connection that the source material does not support. Tencent publishes long-context benchmark results, but a user should test retrieval across the beginning, middle and end of their own documents.

Cost and latency also grow with input length. Sending a million tokens simply because the model accepts them may be wasteful. Retrieval, summarisation and caching can reduce repeated processing. A good production design uses the smallest context that preserves the evidence needed for the answer.

How developers can access Tencent Hy4

There are three practical routes. The first is a Tencent application such as WorkBuddy or CodeBuddy, suitable for users who want the model inside a finished productivity product. The second is a hosted API through Tencent Cloud TokenHub or another provider such as OpenRouter. The third is downloading the weights and operating the model on private infrastructure.

The API path trades control for simplicity. A provider manages GPUs, scaling and model serving, while the customer pays for tokens and accepts the provider’s data terms. Self-hosting offers greater control over data and modifications, but requires accelerator capacity, networking, inference software, monitoring and staff who can keep the service reliable.

For most Indian startups, testing through an API is the realistic first step. A team can measure coding accuracy, multilingual behaviour, latency and cost on its own workload before considering dedicated capacity. Regulated businesses should also review where prompts are processed, how logs are retained and whether sensitive data can be excluded.

Why Tencent Hy4 matters to the open-model race

Hy4 combines three things that are often separated: a very large openly available model, distribution inside mass-market productivity tools, and cloud/API access for developers. That gives Tencent feedback from both end users and technical adopters. The company says it uses a preview-first process so real-world use can inform the official release.

The release also shows how Chinese AI companies are competing beyond price. Long context, coding performance, permissive licences and rapidly integrated applications can attract developers even when a model is expensive to self-host. An open-weight launch can support a provider’s cloud business because many users who are allowed to download the model will still pay someone else to run it.

Readers can compare Hy4 with Lapaas Voice’s coverage of India’s shift toward open-source AI models, Alibaba’s plan to monetise open models for enterprises, and MiniMax’s much larger open-model ambition. The common business model is clear: release weights widely, then compete for usage, hosting and enterprise integration.

What businesses should test before adopting Hy4

Start with a representative task set rather than a public leaderboard. For coding, include real bugs, repository navigation and tests. For office work, include long documents with known answers and deliberately conflicting evidence. For research, score citation accuracy, numerical reasoning and whether the model admits uncertainty.

Measure total cost, not only the listed token price. Long prompts increase input charges and latency; self-hosting adds GPU, power and staffing costs. Test how often the model repeats analysis, because Tencent’s own materials acknowledge that the preview can reason for too long or over-check an answer.

Finally, run security and governance checks. Determine whether tool calls are restricted, prompts and outputs are logged safely, licences and notices are preserved, and the model can be rolled back when the preview changes. Open weights create deployment choice, but the deployer remains responsible for the application.

The primary references are Tencent’s official Hy4 announcement and the official Hy4 repository and licence. Independent analyses including PacketNebula’s benchmark count and AI on Mac’s deployment review provide useful checks on the launch claims.

FAQs

What is Tencent Hy4?

Tencent Hy4 preview is an open-weight mixture-of-experts language model with 770 billion total parameters, about 49 billion active per token and a context window above one million tokens.

Can Tencent Hy4 run on a laptop?

Not in its full standard form. Although only part of the model computes each token, a deployment must make the full weights available. Official examples use multi-GPU datacentre systems; heavily quantised versions still require substantial hardware.

Is Tencent Hy4 free to use commercially?

The official repository uses Apache 2.0, a permissive licence that allows commercial use and modification when its conditions and notices are followed. Infrastructure or hosted API usage still costs money.

Is Tencent Hy4 better than closed AI models?

Tencent reports competitive results, especially in coding and long-context tasks, but most launch benchmarks are vendor-generated. Businesses should run independent tests on their own tasks, cost limits and safety requirements.