Tencent announced the release of Hy4 flash on November 17, 2026, as the newest addition to its Hy model family. The multimodal AI model features 224 billion total parameters and a context length of up to 1M tokens.
Model Architecture and Performance
Tencent stated that Hy4 flash includes 224 billion total parameters, which encompasses 64 billion Engram parameters, alongside 12 billion active parameters. The model supports multimodal inputs and a context length of up to 1M tokens.
According to Tencent, the pre-training incorporates hybrid attention, LatentMoE, and Engram architectural features. By combining optimization using megakernel, 4-bit quantization, and DFly techniques, the model achieves decoding speeds exceeding 150 tokens per second.
Pricing and Availability
Tencent reported that Hy4 flash is priced at approximately US$0.070 per million input tokens, US$0.281 per million output tokens, and US$0.003 per million cached input tokens. The company stated that users can access the model for free on WorkBuddy and receive 50 percent off standard pricing via Tencent Cloud TokenHub from now until November 30.
The model is currently available across Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with OpenRouter integration coming soon.
Key facts and where they come from
- Hy4 flash has 224 billion total parameters and 12 billion active parameters.
Hy4 flash has 224 billion total parameters (including 64 billion Engram parameters) with only 12 billion active parameters
- The model supports multimodal inputs and a context length up to 1M tokens.
supports multimodal inputs and a context length of up to 1M tokens.
- Input pricing is approximately US$0.070 per million input tokens.
priced at approximately US$0.070 per million input tokens
The catch
- The announcement does not provide independent benchmark testing results outside of Tencent's internal evaluations.
- Performance comparisons and per-task cost figures are based on Tencent's internal WorkIndex benchmark.
