WORLDTECH NEWS Global technology intelligence.Contact
AI SINGLE SOURCE

Tencent Launches Hy4 Flash as a Lower-Cost text-and-image Model

Elegant 3D visualization of neural networks showcasing abstract connections in a digital space.
Illustrative photo.Photo by Google DeepMind on Pexels

What happened

Tencent Hy introduced Hy4 flash on November 16, 2026, a natively multimodal (able to work with more than one kind of input, such as text and images) model with a one-million-token context window (how much text a model can consider at once) and substantially lower API (the interface one piece of software uses to talk to another) prices than full Hy4. Tencent is a video game company based in Maling.

The release is aimed at everyday productivity work.

Tencent Cloud’s price list puts Hy4 flash at $0.070 per million input tokens and $0.281 per million output tokens, compared with $0.834 and $2.501 for Hy4. Cached input costs $0.003 per million tokens, against $0.042 for Hy4.

View more Hy4 flash cuts input and output token prices by about 92% and 89% versus full Hy4. Tencent's WorkIndex shifts evaluation toward real office deliverables rather than public benchmark scores.

Tencent positions Hy4 flash for high-volume productivity while full Hy4 keeps the higher intelligence ceiling. Input and output token rates are therefore about 92% and 89% lower, respectively, calculated from those list prices.

Architecture and Serving Optimizations The model contains 224 billion total parameters, including 64 billion Engram parameters, while activating 12 billion parameters per token. Tencent reports that its hybrid-attention, LatentMoE and Engram design reduces the key-value cache size, which stores earlier context computations for reuse, to one twenty-fourth of Hy3’s and one-sixth of Hy4’s.

Key facts

  • Tencent Hy — introduced: Hy4 flash on November 16, 2026, a natively multimodal model with a one-million-token context window and substantially lower API prices than full Hy4
  • Architecture and Serving Optimizations The model — includes: 224 billion total parameters, including 64 billion Engram parameters, while activating 12 billion parameters per token
  • Tencent — reports: that its hybrid-attention, LatentMoE and Engram design reduces the key-value cache size, which stores earlier context computations for reuse, to one twenty-fourth of Hy3’s and one-sixth of Hy4’s

Sources & evidence