Flagship and lightweight weights released together

China's Tencent has released the weights (trained parameters) of "Hy4," the flagship of its "Hunyuan" AI models, and the lightweight "Hy4 flash" in its official repositories on the AI sharing site Hugging Face. On the morning of Nov. 21 Japan time, ORIENS confirmed through Hugging Face's public API and the model cards that both models can be downloaded by anyone. With the weights, companies can run the models on their own servers or further train them on their own data.

Tencent released three versions of each: the standard model, an "FP8" version shrunk to 8-bit data, and a "base" version before additional training. Besides Hugging Face, it also posted them on China's ModelScope and GitCode. Tencent has not said when the weights were published, and ORIENS has not been able to confirm it. According to Hugging Face's update history, the last update to the Hy4 repository was at 12:05 p.m. Nov. 20 Japan time, and to Hy4 flash at 8:25 p.m. Nov. 19 Japan time.

Tencent released the official version of Hy4 through its API on Nov. 4, but did not release the weights at that time. China's IT Home and other outlets reported then that Tencent would release the weights soon. It released Hy4 flash through its API on Nov. 16 and said it would release the weights later (ORIENS reported).

Hy4 also understands video; Hy4 flash handles text and images

According to the model card, Hy4 has 770 billion total parameters, and for each token, a unit roughly equivalent to a word, it activates 49 billion in the text part and 1.23 billion in the part that reads images. It uses a "mixture of experts (MoE)" design that activates only some of many small expert components. It understands video in addition to text and images, and can read 1,048,576 tokens (about 1 million tokens) at once.

Hy4 flash has 220 billion total parameters (including 64 billion in a part called "Engram"), with 12 billion active. It accepts text and image input and also has a 1 million token context window. The Nov. 16 announcement page had put the total at 224 billion parameters.

Running Hy4 requires multiple GPUs. The Transformers documentation says the full model requires multiple GPUs, and the model card shows an example of running the FP8 version split across eight GPUs with the inference software "vLLM." ORIENS has not been able to confirm official guidance on the type or number of GPUs needed.

Transformers support came before the release

On Nov. 17 (11:20 p.m. Nov. 17 Japan time), ahead of the Hy4 release, support for a model called "HY4-VL" was merged into the open-source AI library "Transformers." Transformers is developed mainly by Hugging Face and is used to load many AI models. The first commit in the merged change was titled "add open-source HY4-VL model" and was written on Oct. 2 by a developer using a Tencent email address.

Documentation added at the same time described HY4-VL as combining Hy4's text model with an image-reading component, generating text from images and text, or video and text. Its usage examples used tencent/Hy4 as a tentative model name, noting that it may change before release. The documentation is pre-announcement information found in a public repository. The actual weights were released under the same name, tencent/Hy4.

Support for Hy4 flash was proposed to Transformers on Nov. 18 as a change (PR #50903), but it had not been merged into the main code as of the morning of Nov. 21. The Hy4 flash model card tells users to use the proposed version until it is merged.

License moves from Apache 2.0 to Tencent's own terms with conditions

The August preview, "Hy4 preview," was released with weights under the Apache 2.0 license, which allows broad use including commercial use. The six new models use Tencent's own "Tencent Hy License 1.0." It is based on Apache 2.0 and adds conditions for businesses that offer the model to outside users as Section 10.

Section 10 covers businesses that let third parties use a general-purpose model through APIs and similar means (MaaS). Companies whose worldwide gross revenue exceeds $50 million in any consecutive 12-month period, or uses in products or services with more than 100 million monthly active users, require a separate written agreement with Tencent. Use solely for internal purposes, without making the model or its outputs available to third parties, is exempt.

"Kimi K3.1 Pro," whose weights China's Moonshot AI released on Nov. 4 (Nov. 5 Japan time), also requires a separate agreement from companies with annual revenue above $20 million in the business of offering the model to outside users. Tencent's revenue threshold is 2.5 times that.

Tencent acknowledges gap with top closed models in its own evaluations

In the model card, Tencent called Hy4 one of the strongest open-weight models available today, while acknowledging that a notable gap remains with top proprietary models on challenging internal evaluations such as research-level mathematics, long-horizon coding and advanced visual question answering.

In a chart of 12 evaluations the company published, Hy4 scored 44.4 on the terminal-operation benchmark "Terminal Bench 4.0," below OpenAI's "GPT 6.1 Sol" at 55.5 and Anthropic's "Claude Opus 5.5" at 66.4. On the computer-operation benchmark "OSWorld 2.1," it scored 78.9, close to Claude Opus 5.5's 81.8. All are figures measured by Tencent, and the chart does not state a maximum score.

Hy4 benchmark results published by Tencent, comparing Hy4, Hy4 preview, GLM 5.3, DeepSeek-V4.1-Flash-1017, Kimi K3.1 Pro, Claude Opus 5.5 and GPT 6.1 Sol across 12 evaluations
Hy4 benchmark results published by Tencent, comparing Hy4, Hy4 preview, GLM 5.3, DeepSeek-V4.1-Flash-1017, Kimi K3.1 Pro, Claude Opus 5.5 and GPT 6.1 Sol across 12 evaluations (Source: Hugging Face: Hy4 model card)

In an internal evaluation in which 213 finance professionals scored 579 tasks (out of 4 points), Tencent said Hy4 scored 3.04, above DeepSeek-V4.1-Flash-1017 at 2.97, Kimi K3.1 Pro at 2.68 and "GLM-5.3" at 2.68, ranking first among open-weight models. ORIENS has not found any third-party evaluation.

From announcement to weight release

Tencent announced on Oct. 29 that it would soon release the official version and Hy4 flash, and released Hy4 through its API on Nov. 4 and Hy4 flash on Nov. 16. After the Transformers support was merged, the weights were all on Hugging Face by the morning of Nov. 21 Japan time.

Until the Hy4 series weights were released
  1. Preview "Hy4 preview" released with weights under Apache 2.0Official
  2. Announces official version and lightweight "Hy4 flash" coming soonOfficial
  3. Official Hy4 released through API, weights withheldOfficial
  4. Hy4 flash released through API, weights said to come laterOfficial
  5. HY4-VL support in Transformers, tentative model name `tencent/Hy4`Leak
  6. Hy4 and Hy4 flash weights confirmed public on Hugging FaceOfficial