Tencent Releases Hy4 Weights: 49 Billion of 770 Billion Parameters Active per Token
Tencent has released the weights of its flagship Hy4 model, so it can be downloaded and run on your own servers. It is a mixture-of-experts model with 770 billion parameters, but only 49 billion are active for each token (a fragment of a word), which makes it cheaper to run. In Tencent's own tests on agentic tasks such as fixing bugs in repositories, it performs on par with GPT-6.1 Sol and Claude Opus 5.5, but it is weaker in science. Teams building agents should run it on their own tasks.
AI-processed from TG: @ai_machinelearning_big_data (Machinelearning); edited by Hamidun News
Tencent has delivered on its promise from the presentation and published the weights of its flagship Hy4 model. This means a heavy model at the level of closed flagships can be downloaded, deployed on your own servers, and built into a product without sending data to someone else's cloud. But compared with the August preview, the license has changed, and for some companies this matters more than any benchmark.
What the numbers in the headline mean
Hy4 uses a mixture-of-experts design. Picture a large consulting firm with 770 employees. For each client question, they don't call in everyone at once: they bring in the 49 specialists who are needed. The model works the same way. It has 770 billion parameters in total, but only 49 billion are involved in processing each token (roughly a fragment of a word). That is why it is cheaper and faster to run than a dense model of the same size, where all parameters work at once.
Hy4 has a context window of 1 million tokens. In practice, this means a large code repository or a long archive of contracts fits into a single request. The model understands not only text but also images and video.
It also includes a built-in draft model for speculative decoding. Think of an experienced editor and an intern: the intern quickly drafts several next words, and the editor checks them all at once. If the draft is correct, the answer comes faster, and the quality stays the same.
Alongside the weights, Tencent released a system card of more than 80 pages with safety assessments, including risks in cybersecurity and biosecurity. If model selection has to be approved by a security team, this document will be useful.
License: who now needs permission
The preview version, Hy4 Preview, released in August, was distributed under Apache 2.0, which means it came with almost no restrictions. The final model is released under its own Tencent Hy License.
Hobbyists and researchers can still use the model for free, modify it, and apply it in commercial products. The restriction applies to companies that offer Hy4 as a service. They need a separate written permission from Tencent if either of the following conditions applies:
- the company's annual revenue over any 12 consecutive months exceeds $50 million;
- the product reaches more than 100 million monthly active users in at least one month.
Ordinary deployments, where the model handles individual functions, and internal use within a company do not fall under the restriction. For example, a startup that added Hy4 to its CRM to parse emails, or a corporate team with an internal assistant, can most likely proceed without concern. Cloud providers and large platforms that resell the model as a service, however, will need to negotiate.
What the tests showed and what it costs
An important caveat: Tencent ran most of the comparisons itself, in its own agent harnesses. Independent evaluations may paint a different picture.
According to these measurements, on agentic tasks Hy4 holds its own against the closed GPT-6.1 Sol and Claude Opus 5.5:
- on bug fixing in real repositories on DeepSWE, it scores almost the same as both models;
- on APEX-Agents agentic tasks, it outperforms both;
- on computer control on OSWorld, it beats GPT and falls slightly short of Opus.
On longer and more complex tasks, the results are weaker. On Terminal-Bench 4.0 and FrontierCode, Hy4 trails GPT, and trails Opus by a wider margin. On the HLE exam with tools, it nearly catches GPT, while on image understanding on MMMU-Pro it loses to both. The gap is largest in science: on Terminal-Bench-Science, the model solves nearly three times fewer tasks than GPT and Opus.
Via the API on the Tencent Cloud platform, Hy4 costs $0.834 per million input tokens and $2.5 per million output tokens. A 50% discount applies until December 31.
What to do in practice
If you are building agents for code, interface work, or business processes, run Hy4 on your own tasks. Open weights and a low API price make it a real alternative to closed models, especially when data cannot leave your environment. For scientific and multimodal use cases, check quality separately: that is where the model is weaker.
If you plan to sell access to the model as a service, read the license text first and estimate whether you will hit the revenue or user-count threshold. For small teams the terms remain lenient, but for large platforms, the Apache 2.0 terms from the preview no longer apply.
- Nov 22, 2026Analysts estimate: by 2028, gas will supply 87% of Anthropic's power and coal 11%
- Nov 22, 2026Two AI Judges, 15 LLMs: Judges Favor Their Own Makers by 0.04 Points Out of 1
- Nov 22, 2026OpenAI's New Astra Model Writes 40% Shorter, but Its Hedging Score Fell from 96 to 24
- Nov 22, 2026Imagecaster trains a neural network on your face from 10–100 photos and tries on clothes at home
Key Questions Answered
What the numbers in the headline mean?
Hy4 uses a mixture-of-experts design. Picture a large consulting firm with 770 employees. For each client question, they don't call in everyone at once: they bring in the 49 specialists who are needed. The model works the same way. It has 770 billion parameters in total, but only 49 billion are involved in processing each token (roughly a fragment of a word). That is why it is cheaper and faster to run than a dense model of the same size, where all parameters work at once.
What are the key details on License: who now needs permission?
The preview version, Hy4 Preview, released in August, was distributed under Apache 2.0, which means it came with almost no restrictions. The final model is released under its own Tencent Hy License. Hobbyists and researchers can still use the model for free, modify it, and apply it in commercial products. The restriction applies to companies that offer Hy4 as a service.
What the tests showed and what it costs?
An important caveat: Tencent ran most of the comparisons itself, in its own agent harnesses. Independent evaluations may paint a different picture. According to these measurements, on agentic tasks Hy4 holds its own against the closed GPT-6.1 Sol and Claude Opus 5.5: on bug fixing in real repositories on DeepSWE, it scores almost the same as both models; on APEX-Agents agentic tasks, it outperforms both; on computer control on OSWorld, it beats GPT and falls slightly short of Opus.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.