Tencent's Hy team presents plans for larger models and a 24GB deployment roadmap
Read more
Pandaily
pandaily.com

Tencent's Hy team presents plans for larger models and a 24GB deployment roadmap

The Hy team from Tencent presented its intentions regarding the development of larger models and smaller local deployment options during a Q&A session on the public Reddit forum. This discussion took place early on November 21st in Shanghai, following the release of the flagship Hy4 weights the previous day.

According to a Chinese summary of responses prepared by Zhidx, the team stated that the next Hy model will be significantly larger, and they will continue to invest in large-scale reinforcement learning training. It is important to note that these are development intentions, not an announcement of a new checkpoint, confirmed parameter count, or a set release date.

Regarding smaller models, the team described Flash-Lite as planned work and discussed models intended for devices with 24GB of GPU memory. The roadmap also featured Hy4-mini, and a dense model with parameters ranging from 24 to 31 billion was under consideration. The responses do not guarantee that all these products form a single entity or that any proposed configuration is currently available.

Hardware requirements demand similar attention. The mention of 24GB of video memory itself does not define the total system RAM, quantization, context length, or achievable speed. These details are necessary before developers can consider a roadmap item as a specification for consumer hardware ready for implementation.

The architectural discussion covered trade-offs between Grouped Query Attention (GQA) and Multi-Head Latent Attention (MLA). As summarized by Zhidx, the team noted that MLA provides better model quality at a comparable key-value cache size, but its computational intensity complicated decoding. The implementation of the Sparse Attention DSA mechanism reduced this load and changed the team's preferred balance.

The team also cautioned against assuming that greater sparsity in the Mixture-of-Experts architecture always improves deployment. An overly sparse model might consume significant GPU memory without fully utilizing available computational power; reducing active parameters while maintaining overall size can also negatively affect quality. These statements are design explanations, not universal benchmarking results for every implementation.

The initial open AMA post confirms the session and participation of individuals, but the details of the answers are based on a Chinese compilation, not independently obtained transcripts of comments. The practical value lies in a clearer presentation of research and engineering priorities, rather than another announcement that Hy4 weights are available for download.

Similar stories

Tencent releases Hy4 model weights with a million-token context window
Read more
technode.com

Tencent releases Hy4 model weights with a million-token context window

Tencent announced on Friday the weights of its flagship artificial intelligence model, Hy4, allowing developers to download and implement the model independently. This release follows the initial launch of Hy4 on November 4th and makes the flagship model available under the Apache 2.0 license, which permits both commercial use and modification.

The Hy4 model is capable of understanding text, images, and video, featuring a context window of approximately one million tokens. Its 'Mixture of Experts' language framework contains 770 billion parameters, with 49 billion parameters activated for processing each token.

According to Tencent's statement, this version of the model surpasses the Hy4 preview in areas such as instruction following, long-context understanding, advanced task execution, and scientific research.

The model supports workflows covering graphical user interfaces and command-line tools, as well as frontend development, game design, and office tasks. Tencent provides deployment instructions for vLLM and SGLang, along with resources for fine-tuning and a safety evaluation map. The release of the flagship model followed the separate unveiling of the smaller Hy4 flash model.

Tencent's Hunyuan Hy4 flash model exceeded 50 trillion tokens in its first day of use
Read more
technode.com

Tencent's Hunyuan Hy4 flash model exceeded 50 trillion tokens in its first day of use

Tencent announced that its Hunyuan Hy4 flash model processed over 50 trillion tokens in just the first day after its launch on November 16. This figure was the highest among all models in the Hunyuan series for their first day of operation.

The strong start of Hy4 flash also reflected in its performance across various platforms. On November 17, the model ranked first in the daily LLM rating on OpenRouter.

Furthermore, Tencent noted that on its WorkBuddy platform, Hy4 flash demonstrated ten times greater usage compared to its predecessor, Hy4 preview, and helped complete over 50 million real user tasks.

According to Tencent data, Hy4 flash utilizes a hybrid activation architecture and has a total of 224 billion parameters, including 64 billion Engram parameters, with 12 billion parameters activated during inference. The model also supports multimodal processing and a context window of up to 1 million tokens, achieving a generation speed of about 200 tokens per second.

Of particular interest is the balance between performance and cost offered by Hy4 flash. According to Tencent's pricing, the model costs only 0.5 yuan ($0.07) per million input tokens and 2 yuan ($0.3) per million output tokens, while cached input tokens cost only 0.02 yuan ($0.003) per million.

Based on Tencent's internal WorkIndex benchmark, which assesses standard office tasks, Hy4 flash costs only one-thirteenth compared to Hy4 preview and one-twentieth compared to Hy4 for completing a single task. Its cost is also approximately one-eighth of the cost of mainstream lightweight models in the same category.

These advantages make the model particularly suitable for frequently performed office operations such as creating and editing presentations, document processing, spreadsheet recovery, audio and video processing, and visual understanding.

The launch followed a previous surge in demand for Hy4. When the model first appeared on WorkBuddy on November 4, demand driven by its agent capabilities exceeded expectations, and many users reached usage limits on the very first day. In response, Tencent increased Hy4 inference clusters tenfold and expanded free usage quotas.

However, this time the team apparently prepared more thoroughly for high competition before the launch of Hy4 flash, and there are currently no reports of widespread usage limit issues.

This development also reflects a broader shift in the competitive landscape of the artificial intelligence industry. As daily token usage in China continues to grow, tokens are becoming an increasingly important metric for measuring AI adoption. Meanwhile, local authorities across the country are introducing new measures, ranging from computing power and model vouchers to token-based funding, to make AI services more accessible.

Competition is now determined not only by the number of parameters or model performance. Achieving strong results while maintaining low operating costs is becoming equally important. The debut of Hy4 flash with 50 trillion tokens sends a clear signal: the market needs not only more powerful AI models but also models that are accessible enough for frequent daily use.

Nevertheless, there remains a gap between high initial usage and long-term commercial value. Competition among lightweight AI models may increasingly boil down not only to prices and token usage. The real challenge is to provide solutions for real business needs stably and reliably while keeping overall costs low. For Tencent's Hunyuan, the strong debut of Hy4 flash is a promising start, but its ability to convert initial usage momentum into sustainable user value will be a key test in the future.

Tencent releases Hy4 flash model weights for local deployment
Read more
technode.com

Tencent releases Hy4 flash model weights for local deployment

Tencent announced the weights of its artificial intelligence model, Hy4 flash, on Wednesday. This release came two days after the model was introduced to users via the company's applications and cloud API. Developers can now download the standard version of the model, its FP8 format version, as well as the base model for further training.

This release aligns with Tencent's previous statement regarding plans to make the model weights publicly available. The Hy4 flash model is capable of processing both text and image input, and its context window reaches one million tokens.

The model has approximately 224 billion parameters, with about 12 billion parameters activated per token. The main goal of Hy4 flash is to improve daily performance in tasks such as document preparation, research, and spreadsheet work. Tencent accompanied the weights with relevant deployment materials, distributing them under the Tencent Hy license.

Popular