Tencent opens Hy4 weights, a 770B model trained for computer use

The multimodal mixture-of-experts model activates 49 billion parameters per token, reads up to 1 million tokens and was trained to drive desktop software and command-line tools.

Illustration of a robot built from purple and teal cubes, a few of them glowing, reaching toward floating screens that show a 3D design program and a command-line terminal. Image: AgenticFeed illustration

Tencent released the weights for Hy4, its flagship AI model, on November 20, letting developers download and run a model the company says was trained to operate software on a computer. Hy4 has 770 billion total parameters, with 49 billion active for each token, and a context window of 1,048,576 tokens, according to the model card Tencent's Hy team published on Hugging Face.

Tencent first launched Hy4 on November 4 through its own cloud and third-party platforms including OpenRouter, and said at the time that it planned to publish the weights and system cards later. Sulat reported that the API launched at $0.834 per million input tokens and $2.501 per million output tokens.

Built for agent work

The model card describes several agent uses. Hy4 "supports computer-use workflows across command-line interfaces (CLI) and graphical user interfaces (GUI)," the Hy team wrote, adding that targeted training on specialized software such as Blender and CAD tools, along with long-horizon workflows that span multiple applications, improves planning and execution.

Tencent also pitches Hy4 as a long-running personal assistant. Given work context, the model can draw on chat history, emails, documents, calendars and meeting notes to track projects, surface action items, prepare meeting briefs and flag deadlines, according to the card.

For finance work, Tencent said it built thousands of simulated workspaces and tens of thousands of tasks. It tested Hy4 with 213 finance professionals, including 153 experts from financial institutions, across 579 real-world tasks. Hy4 averaged 3.04 out of 4, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each, Tencent reported.

The company said targeted search upgrades cut Hy4's hallucination rate in internal evaluations from 15.76% on the earlier Hy4 preview to 5.45%, and reduced speculative errors from 23.1% to 5.6%. Paired with Tencent's Hyra 2.0 harness, the model optimized internal pre-training infrastructure and autonomously trained a smaller variant that beat an internal baseline on multiple benchmarks, the card said. These results come from Tencent's own tests and have not been independently reproduced.

Tencent was candid about limits. "A notable gap remains between it and top proprietary models," the card said, citing research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering.

What is in the release

The language backbone has 78 layers, according to the model card: one dense layer and 77 mixture-of-experts layers, each with 256 routed experts and one shared expert. A router selects eight routed experts per token. A 1.23 billion parameter vision encoder lets the model read images and video, and a 3.3 billion parameter drafter called DFly2 speeds up generation through speculative decoding.

Tencent posted the weights on Hugging Face and other hubs including ModelScope. The instruct model comes with an FP8 quantized copy, and the pre-trained base model is also available. Sulat reported that the instruct model's files alone total about 1.55TB.

Prebuilt images are available for the vLLM and SGLang serving engines, according to the model card, and Tencent's own vLLM example runs the FP8 model across eight GPUs. A safety system card came out alongside the weights.

Licence terms

Hy4 ships under the Tencent Hy License, according to the model card. The earlier Hy4 preview, released in August, used Apache 2.0, Sulat reported, and the new licence requires a separate agreement for model-as-a-service businesses above $50 million in revenue or products with more than 100 million monthly users. TechNode described the release as Apache 2.0, which does not match the licence listed on the model card.

Tencent says Hy4 already runs in its products, including CodeBuddy, WorkBuddy, Yuanbao and the AI assistant in its WeCom work app, with integration planned for QQ and Tencent News.

Source: Hugging Face (Tencent), Tencent Hy Team, “tencent/Hy4 model card”

Additional sources: Sulat, JP Caparas, “Tencent releases open weights for Hy4, a 770B model under its own licence”; Tencent, “Tencent Launches Hy4, Joining the Ranks of the World's Leading Open-Source AI Models”; TechNode, “Tencent releases Hy4 model weights with million-token context”

Get the AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam, and you can unsubscribe at any time.

Want to reach readers following AI agents? Advertise with AgenticFeed

Next event · 30 Nov, 2026

AWS re:Invent 2026 in Las Vegas

AWS’s annual conference runs Nov. 30 to Dec. 4. Matt Garman’s CEO keynote is on Dec. 1, and the catalog lists more than 2,200 sessions.

The AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam.