AI Aulis

AI NEWS IN ENGLISH · 19 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

38 stories · next issue tomorrow morning · How the paper is made →

← Issue 19 November 2026Paper

Tencent’s Hy4 flash weights available for download – standard version takes up about 440 gigabytes

AI-generated contentWritten by AI

github.com, api.github.com, hy.tencent.ai, pandaily.com, arXiv · 5 sources · Sources published: 23 October–18 November 2026 · Some source publication dates are unknown

The weights of Tencent’s Hy4 flash model can now be downloaded to your own hardware: the standard version takes up about 440 gigabytes, even though only about 12 billion parameters take part in computation per token. The licence requires the largest service providers to obtain separate permission.

Pandaily reported on 18 November, citing IT Home, that Tencent has released the weights of Hy4 flash. Tencent launched the model through its API two days earlier. Three versions are available for download: an instruction-tuned model, a version of it quantised to FP8, and a pretrained base model intended for further training.

Hy4 flash is a mixture-of-experts (MoE) model, so only part of the model’s networks process each token. Tencent’s GitHub page says that in each MoE layer a token activates 16 of 512 routed experts plus one shared expert used by all tokens. About 12 billion parameters take part in computation per token, out of 220 billion in total. The company’s website gives the total as 224 billion.

The number of active parameters reflects how much computation the model needs, but all the weights still have to be stored. Pandaily reported that the standard version’s weights take up about 440.5 gigabytes and the FP8 version’s about 287.1 gigabytes. The publication said these figures are not minimum GPU memory requirements and that quantisation does not make the model suitable for consumer hardware. Pandaily confirmed that the download can be started, but it did not download the full weights or run the model.

Tencent’s deployment instructions run the model on eight GPUs. Pandaily said the guide for the vLLM serving software starts the model with a context of 8,192 tokens. A 1 million-token context window therefore does not guarantee that every hardware setup can handle a context that long.

Pandaily noted that the Tencent Hy License 1.0 is not an Apache 2.0 licence. Separate written permission is required if a company offering the model as a service exceeds $50 million in worldwide revenue over 12 consecutive months, or if the model is used in a product or service with more than 100 million monthly users. Pandaily describes Hy4 flash as an open-weight model on Tencent’s terms.

The performance figures are Tencent’s own. In the company’s internal WorkIndex benchmark, Hy4 flash scored 42.1 points and Hy4 preview 39.7. Tencent said Hy4 flash’s average cost per task was 1/13 of Hy4 preview’s. For complex tasks such as long software development tasks, the company recommends the more capable Hy4.

The weights are available from Hugging Face, among other places. Through the API, the model costs $0.07 per million input tokens and $0.281 per million output tokens. Tencent Cloud TokenHub offers a 50 per cent discount on API prices until 30 November.

What this is based on

  1. github.com

    Publisher: github.com · published 18 November 2026 · retrieved 19 November 2026 at 05:32 (UTC+02:00) · discovery channel: GitHub

  2. 腾讯混元 Tencent Hunyuan (GitHub)

    腾讯混元 Tencent Hunyuan (GitHub) · Publisher: api.github.com · source publication time unknown · retrieved 19 November 2026 at 04:02 (UTC+02:00) · discovery channel: web

  3. Tencent Hy

    Publisher: hy.tencent.ai · published 16 November 2026 at 14:12 (UTC+02:00) · retrieved 17 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

  4. Tencent Releases Downloadable Hy4 flash Weights in Three Variants

    contact@pandaily.com (Pandaily) · Publisher: pandaily.com · published 18 November 2026 at 09:10 (UTC+02:00) · retrieved 19 November 2026 at 04:22 (UTC+02:00) · discovery channel: RSS feed

  5. HyScale: Scaling Long-Context Training for Learned Sparse Attention

    Deng, Yangshen, Huang, Xinjing, Wang, Yulong, Liu, Yunzhuo, Hu, Jun, Tan, Shuyong, Tseng, Chun-Chih, Han, Mingji, Ying, Ray, Zhang, Key, Mai, Luo · Publisher: arxiv.org · published 23 October 2026 at 04:54 (UTC+03:00) · This version published: 4 November 2026 at 20:23 (UTC+02:00) · retrieved 19 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

Report an error in this story