📣 Send us your press release
Site updates every 15 minutes
Technology

Tencent introduces Hy4 flash AI model focused on cost efficiency

Tencent has launched its Hy4 flash language model, designed to optimize performance-to-cost ratios for enterprise and daily tasks.

16 November 2026

Tencent has officially unveiled the Hy4 flash model, which features a total parameter count of 224 billion with 12 billion active parameters per token. The model supports multimodal inputs and a context length of one million tokens, with the company emphasizing a significant reduction in API costs compared to its predecessor, the Hy4.

To achieve this efficiency, the engineering team implemented architectural adjustments, including LatentMoE and Engram parameter modules, to reduce KV cache size. During inference, the model utilizes 4-bit quantization and specialized kernel operations to maintain decoding speeds exceeding 150 tokens per second.

Internal testing across various work-related benchmarks indicates that Hy4 flash provides performance levels comparable to more resource-intensive models while maintaining lower operational expenses. The model is currently available across Tencent's service portfolio, with plans for a future open-source release.

Original source: ithome.com