Key Points
- Tencent HuanYuan has open-sourced Hy4 Flash, Hy4 Flash-FP8 and Hy4 Flash Base, just two days after the model’s launch.
- Hy4 Flash contains 224 billion parameters, with 12 billion activated during each inference. The FP8 version is designed to reduce GPU memory requirements, although deployment remains resource-intensive.
- Developers can use and modify the model under the Tencent HuanYuan License, while large API providers meeting specified revenue or user thresholds must obtain additional authorization.
Three Versions Target Different Development Needs
Tencent HuanYuan has released the weights for Hy4 Flash across three versions, giving developers different options for deployment and customization.
Hy4 Flash is the instruction-tuned production model, designed for direct use in dialogue, reasoning and agent-related tasks. Hy4 Flash-FP8 is an 8-bit quantized version intended to reduce video memory requirements during deployment. Hy4 Flash Base, meanwhile, is the pre-trained model without instruction tuning, allowing developers to continue training and adapt it for specific applications.
The model has 224 billion parameters, with 12 billion activated per inference. Despite the memory savings offered by FP8 quantization, the model still requires substantial computing resources, making deployment a significant consideration for developers.
Availability and Community Support
Tencent has made the model available through GitHub, Hugging Face, ModelScope and GitCode. Community developers have also produced a GGUF-quantized version, which is intended to make the model easier to run across a wider range of hardware environments.
The availability of multiple weight formats and community-supported quantization options gives developers additional flexibility when evaluating deployment requirements and adapting the model to different technical environments.
Licensing and the Hy4 Flagship Model
The release is governed by the Tencent HuanYuan License, which permits developers to use and modify the model weights under specified terms. However, the license includes additional conditions for large-scale model API providers. Companies reaching certain revenue or user-scale thresholds must obtain separate authorization.
Tencent’s larger Hy4 flagship model, which contains 770 billion parameters, has not yet been fully open-sourced. The company says it is accelerating the release process, but has not provided a specific timeline.
What Developers Should Watch Next
The next developments will include the progress of the 770-billion-parameter Hy4 release and how developers adopt the three Hy4 Flash versions. Deployment costs, hardware requirements and the license’s authorization thresholds will remain important considerations for organizations planning commercial applications.
Comparison, examination, and analysis between investment houses
Leave your details, and an expert from our team will get back to you as soon as possible
* This article, in whole or in part, does not contain any promise of investment returns, nor does it constitute professional advice to make investments in any particular field.
To read more about the full disclaimer, click here- Lior mor
- •
- 7 Min Read
- •
- ago 6 hours
SKN | Waymo Pauses U.S. Freeway Robotaxi Rides After California Safety Ban: Will Regulatory Friction Slow Autonomous Commercialization?
Alphabet Inc.’s autonomous driving unit Waymo halted driverless robotaxi rides on freeways across the United States on Tuesday, November 17,
- ago 6 hours
- •
- 7 Min Read
Alphabet Inc.’s autonomous driving unit Waymo halted driverless robotaxi rides on freeways across the United States on Tuesday, November 17,
- Ronny Mor
- •
- 7 Min Read
- •
- ago 6 hours
SKN | Nvidia Board Warned of “Irreplaceable” Talent Loss in Employee Letter: Can Directors Hold the Line Ahead of Q3 Earnings?
Nvidia Corp.’s board of directors has received a direct warning from senior internal engineering leadership that top technical talent is
- ago 6 hours
- •
- 7 Min Read
Nvidia Corp.’s board of directors has received a direct warning from senior internal engineering leadership that top technical talent is
- Lior mor
- •
- 6 Min Read
- •
- ago 10 hours
SKN | Apple Works to Restore Services After Internal Update Triggers Outage
Apple is working to restore remaining systems after an internal update disrupted several customer-facing services, including iCloud and Apple
- ago 10 hours
- •
- 6 Min Read
Apple is working to restore remaining systems after an internal update disrupted several customer-facing services, including iCloud and Apple
- sagi habasov
- •
- 5 Min Read
- •
- ago 11 hours
SKN | Skydance Restructures Film Leadership at Paramount and Warner Bros
Skydance Establishes Studio Leadership Structure Skydance has announced a series of leadership appointments across Paramount Pictures and Warner Bros Pictures,
- ago 11 hours
- •
- 5 Min Read
Skydance Establishes Studio Leadership Structure Skydance has announced a series of leadership appointments across Paramount Pictures and Warner Bros Pictures,