Back to Blog
    AI & AutomationNov 21, 2026

    Tencent Hy4 Open Weights: A 770B AI Model You Can Self-Host

    Amir

    Amir Farid

    7 min read

    Tencent Hy4 Open Weights: A 770B AI Model You Can Self-Host

    On 20 November 2026 Tencent published the downloadable weights for Hy4, its 770-billion-parameter flagship AI model, on Hugging Face, ModelScope and GitCode. Any team with serious GPU hardware can now run it on its own servers: no API key, no per-token bill, and no customer data leaving the building. The release comes in three versions — the Hy4 instruct model, an FP8-quantised version that needs less memory, and the raw pre-trained base model. It completes the open-weight promise Tencent made when Hy4 launched as an API on 4 November. This post explains what Hy4 is, what Tencent claims it can do, what it costs to run, and when your business should self-host it.

    What Tencent released on 20 November

    The news line on the official model card reads: "We open-source Hy4, Hy4-FP8, and the pre-trained Hy4 base model weights." The three downloads are on the Hy4 model card on Hugging Face, mirrored on ModelScope and GitCode. Hy4 is the instruction-tuned model for chat and agent work. Hy4-FP8 is the same model compressed to 8-bit floating point, so it runs on less memory. Hy4 base is the raw pre-trained model, for research and custom fine-tuning.

    This is a distinct event from the launch: we covered the Hy4 API launch on 4 November, when the model was only available through Tencent's hosted API. The weights release is the part that matters for businesses that want control. It continues a busy month for open-weight releases, with Mistral Large 4's open weights, Qwen4 Max and Meituan's LongCat-2.5 all landing in recent weeks.

    Inside Hy4: the numbers that matter

    A GPU server cluster with an open rack in a dark data centre

    The language backbone has 770 billion total parameters, but only 49 billion are activated per token, plus 1.23 billion for vision. It has 78 layers — one dense layer and 77 mixture-of-experts layers — each holding 256 routed experts plus one shared expert, with the top eight routed experts activated for every token.

    A 1M-token context window

    Hy4 supports a 1-million-token context window (1,048,576 tokens) and understands text, images and video. Its Gated DeepSeek Sparse Attention with IndexCache keeps long-context work affordable. Built in is DFly2, a 5-layer, 3.3-billion-parameter drafter that reuses Hy4's KV cache to predict several tokens ahead without extra memory.

    What Hy4 can actually do for a business

    Tencent describes Hy4 as a model built for long-horizon work. The capabilities below are Tencent's claims, not independent verification.

    Computer use: it can operate software

    Hy4 supports computer-use workflows across command-line and graphical interfaces, with targeted training on specialist software such as Blender and CAD and on long-horizon workflows spanning multiple applications. That is the capability that matters most for AI agents that do real work: opening the right tool, clicking through a sequence and carrying a task across apps. It is the same direction as DeepSeek's open-source SuperAgent and the wider AI automation shift.

    Finance and office work

    For finance and office work, Tencent built thousands of realistic workspaces and tens of thousands of tasks. Across 579 real-world finance tasks judged with 213 finance professionals, Hy4 scored 3.04 out of 4, ahead of DeepSeek-V4.1-Flash at 2.97 — first among open-weight models, according to Tencent. Its internal evaluations report the hallucination rate fell from 15.76% to 5.45% versus the Hy4 preview.

    Frontend, games and the personal assistant

    For frontend work, Hy4 is tuned for complex UI interactions, business logic and full-stack integration. For games it is adapted for Unreal Engine, Unity and Godot, with Tencent reporting gains of 19.1% on GameDev and 9.8% on GameLogic versus the Hy4 preview. On TenureTrail, its internal benchmark of 123 research tasks, Tencent says Hy4 ranks first among open-weight models. If your team builds products, our custom software development and SaaS development services turn a model like this into a working product.

    The benchmarks — and why to read them carefully

    Tencent's few-shot scores for the Hy4 base model, against its own Hy4 flash, DeepSeek-V4 and Kimi K3.1 Pro in the same table, are MMLU 91.6, MMLU-Pro 73.9, SuperGPQA 62.7, SimpleQA 57.2 and MMMU-Pro 59.6, with Hy4 base highest on most rows.

    Caveats apply. They come from Tencent's own evaluation suite, on the base model rather than the instruct model you would deploy, and Tencent notes the base model saw large amounts of synthetic instruction data in mid-training, which can distort few-shot scores. Benchmarks signal a model is in the top tier; they do not prove how it will perform on your work. That is also why our guide to AI agent development costs in the UK starts with a pilot, not a purchase.

    What it costs to run Hy4 on your own servers

    The hardware reality

    Tencent's official vLLM recipe runs Hy4-FP8 with tensor-parallel size 8 — the workload split across eight GPUs; the SGLang recipe uses the same eight-way split. Even the FP8 version needs a multi-GPU server or a cloud GPU cluster. Without existing GPU infrastructure, self-hosting means buying servers or renting cloud GPUs by the hour, plus engineering time to keep it healthy — an ongoing cost, not a one-off download. Our custom software development cost guide for the UK explains how infrastructure costs like these are usually estimated.

    Tencent recommends vLLM or SGLang for production serving, with official container images and recipes, a GGUF build for testing, and recommended settings of temperature 0.9 and top_p 1.0. Tencent also provides a complete fine-tuning pipeline.

    Hosted API vs self-hosted

    • Self-hosted: your data stays in-house, no per-token bill, full control over fine-tuning — but you pay for hardware, power and an engineer.
    • Hosted API: zero infrastructure, pay only for what you use — but your data leaves your network and costs scale with every token.
    • Rule of thumb: steady, high-volume usage or data that cannot leave your premises favours self-hosting; spiky or experimental usage favours the API.

    Most businesses should prototype on the hosted API first — the same staged approach we recommend for AI agents for small business automation and for RAG chatbots that read private company documents.

    The licence and the fine print

    Hy4 is released under the Tencent Hy License — that is what the model card states, with a pointer to the LICENSE file for details. The card does not spell out the commercial terms, so read the LICENSE file before using the model commercially, and get legal advice if your use case involves redistribution.

    Tencent is also candid about the model's limits: a notable gap remains versus the top proprietary models, particularly on research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering. If your work is research-grade maths or very long coding tasks, Hy4 is not yet the answer.

    Should your business self-host Hy4?

    Self-host Hy4 if your data cannot leave your premises, if your AI usage is steady and high enough that per-token bills hurt, or if you plan to fine-tune a model on your own data.

    Stick with an API if your usage is unpredictable, your team has no infrastructure experience, or you need the best available reasoning right now. A model is only one part of an agent: it still needs tools, data connections and guardrails before it does useful work. Our AI agent development team builds that whole layer, and our chatbot vs AI agent guide explains the difference. If you want help scoping a self-hosted deployment, talk to us — a short conversation usually settles whether the hardware spend is worth it. Our AI calling agent page shows how a voice agent fits the same stack.

    Frequently asked questions

    Is Tencent Hy4 really free to download?

    The weights are downloadable from Hugging Face, ModelScope and GitCode at no charge: Hy4, Hy4-FP8 and Hy4 base. But "free to download" is not "free to run" — you still need multi-GPU hardware.

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

    No. Tencent's serving recipes split the model across eight GPUs, and even the FP8 version needs a multi-GPU setup. Production use means server-grade GPU hardware or cloud GPU rental.

    How does Hy4 compare with other open-weight models?

    On Tencent's own table, Hy4 base leads DeepSeek-V4 and Kimi K3.1 Pro on most rows, including MMLU at 91.6 and MMLU-Pro at 73.9 — vendor-reported scores.

    Should my business self-host Hy4 or use the hosted API?

    Self-host if your data must stay in-house, your usage is steady and high-volume, or you plan to fine-tune. Use the hosted API if your usage is spiky or small.

    #Tencent Hy4#open-weight AI#self-hosted AI#computer use#AI agents