Selectel has added the Hy4 flash model to its AI router, which provides access to models via an OpenAI-compatible API. The model was introduced by Tencent on November 16 and supports both text and image inputs with a context window of one million tokens.

Hy4 flash features 224 billion parameters, with 12 billion active per token. It incorporates a 64 billion parameter Engram memory, based on a lookup table concept from DeepSeek. The computational component accounts for approximately 160 billion parameters.

The architecture utilizes hybrid attention and LatentMoE, resulting in a KV cache significantly smaller than previous versions. Tencent estimates the decoding speed exceeds 150 tokens per second. The model is positioned for daily tasks like document processing and basic frontend development.

In internal blind tests, Hy4 flash performed comparably to Claude Haiku 5.5 and outperformed GLM-5.3-Flash, while being 18 times cheaper than Haiku. However, it trails behind the full Hy4 model in complex programming tasks, scoring 19.2 points in Terminal-Bench 4.0 compared to 44.4.