Tencent made three Hy4 Flash open-weight variants available on Hugging Face, alongside hosted access, a 1M-token context window and license terms.

Tencent opens Hy4 Flash weights on Hugging Face

Tencent opens Hy4 Flash weights on Hugging Face

Image: AI-generated editorial illustration

Tencent said Hy4 Flash was moving from hosted access into open weights after its Nov. 16 launch, broadening developer access to the multimodal model. The company announced the model across Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, then later added downloadable variants on Hugging Face and related repositories.

Tencent made three Hy4 Flash open-weight variants available after its Nov. 16 launch, broadening access to the multimodal model beyond its hosted products and API channels. The company said the model is available across Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, and that OpenRouter integration was coming soon in the original announcement. Tencent’s Hugging Face repository later listed Hy4 Flash, Hy4 Flash-FP8 and Hy4 Flash-base.

What did Tencent release in open weights?

Tencent published three open-weight variants of Hy4 Flash: the main instruction model, Hy4 Flash-FP8 and Hy4 Flash-base. The company said Hy4 Flash-base is a pre-trained base model rather than an instruction-tuned chat model, while Hy4 Flash and Hy4 Flash-FP8 are instruct models.

Tencent’s GitHub repository says the company is offering the model for direct use and also providing serving and fine-tuning code. The repository points developers to vLLM or SGLang for production serving, notes that Transformers support had not yet merged into the main branch, and says API users can expose Hy4 Flash through an OpenAI-compatible interface after deploying a vLLM server. Tencent also said it provides AngelSlim as a compression toolkit and a GGUF compressed version for end-to-end testing.

How large is Hy4 Flash and how does it work?

Tencent’s official Hugging Face model card describes Hy4 Flash as a multimodal model with 220 billion total parameters, 64 billion of them Engram parameters, and up to a 1 million-token context window. The model card says it accepts text and image inputs and generates text output.

The model card says Hy4 Flash uses a 42-layer language backbone with one dense layer and 41 MoE layers. Tencent describes three architectural changes versus Hy4: hybrid attention, LatentMoE and an n-gram embedding system called Engram. Its property table lists 155.5 billion parameters excluding Engram, ViT and DFly2, plus 63.5 billion Engram parameters, a 0.43 billion-parameter vision encoder and a 0.85 billion-parameter drafter model called DFly2.

Tencent says the model’s vision encoder is adapted from SigLIP2-So400m-Patch16-Res512 and connected through an MLP projector. The property table also lists a 120,832-token vocabulary, identity Hyper-Connections with four residual streams, a 512-token causal sliding-window attention window and sparse context selection of up to 2,048 tokens.

What is Tencent claiming about speed and cost?

Tencent positions Hy4 Flash for everyday productivity tasks and recommends the larger Hy4 for long-horizon software engineering. The company said the model is designed for deep search, document processing, slide creation, simple frontend demos and computer use.

On performance, Tencent said its pre-training reduced KV cache size to one 24th of Hy3 and one sixth of Hy4, reinforcement learning cut average end-to-end output tokens by 50%, and inference optimizations push decoding speed past 150 tokens per second. Tencent also said its internal WorkIndex benchmark scored Hy4 Flash at 42.1 versus 39.7 for Hy4 Preview, with average cost per task at one 13th of Hy4 Preview, one 20th of Hy4 and one eighth of DeepSeek-V4.1-Flash-1017 Max tier peak-hour pricing.

Tencent said the blind evaluation behind that comparison involved 85 internal WorkBuddy users, 259 tasks and 27 categories. It reported user satisfaction of 2.79 out of 4 versus Claude Haiku 5.5 at 2.81 out of 4, with a median completion time of 6.7 minutes and cost of RMB 0.26 per task. The company also said Hy4 Flash recorded the lowest cost per task and shortest median completion time among the evaluated models.

How can developers use Hy4 Flash now?

Tencent says Hy4 Flash is available across its own product stack, including TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis. The company disclosed list API pricing of about $0.070 per million input tokens, $0.281 per million output tokens and $0.003 per million cached input tokens, and said users could access Hy4 Flash for free on WorkBuddy and receive 50% off standard pricing through TokenHub through Nov. 30.

In a Nov. 19 post, Plotly reported that Hy4 Flash was also available through Tencent’s API and OpenRouter. Plotly separately reported that Hy4 Flash passed 39 of 43 questions on its PlotlyDataAnalyticsBench for $0.13, with a median of 29 seconds per question.

Tencent’s license is not Apache 2.0, even though it is derived from Apache License 2.0. The license file says it is a separate Tencent Hy License, Version 1.0, not endorsed by or affiliated with the Apache Software Foundation, and it adds special commercial terms for large model-as-a-service businesses in Section 10. The license covers model weights, parameters, configuration files, inference and training code, and documentation, including fine-tuned, quantized, distilled, merged, pruned or otherwise modified versions of the model.

Sources