~/AI MODELS/tencent-releases-hy4-flash-model-with-high-cost-efficiency

Tencent Releases Hy4 Flash Model With High Cost Efficiency

Tencent released Hy4 flash, a multimodal large language model featuring 224B total parameters, 12B active parameters per token, and a 1M token context length. The model dramatically reduces single-task inference costs to about 5% of the standard Hy4 model through co-optimized training and inference. This release marks a significant breakthrough in large language model efficiency, achieving high daily work performance while offering vastly lower API pricing and operating expenses. By optimizing memory bottlenecks and generation speed, it enables broader commercial deployment for complex daily tasks. Hy4 flash incorporates structural innovations like LatentMoE and Engram parameters, cuts KV cache size significantly, and employs megakernel, 4bit quantization, and dflash to achieve a decoding speed exceeding 150 tokens per second. It is currently deployed across multiple Tencent products and is slated for upcoming availability on OpenRouter.

## BACKGROUND

Large language models often face high computational and memory bottlenecks during inference, which Mixture-of-Experts (MoE) architectures and KV cache optimizations help mitigate. Techniques like LatentMoE route tokens through a compressed latent space to maximize accuracy per FLOP, while Engram modules use scalable lookup-based memory to expand capacity efficiently.

## REFERENCES

## KEYWORDS

#AI Models#Large Language Models#Tencent#Model Optimization#MoE

$ subscribe --daily