Tencent’s new Hy4 Flash is ridiculously cheap, but there’s a catch

Tencent’s new Hy4 Flash packs 224 billion parameters and costs just $0.07 per million input tokens. In the company’s own tests, it came close to Claude Haiku 5.5 while costing 18 times less per task. But there’s a noticeable performance tradeoff.

Tencent Hy logo against gradient orange background

Tencent has released Hy4 Flash, a smaller version of its recently launched Hy4 model. It comes with 224 billion parameters and a one-million-token context window, but what makes it interesting is how little Tencent is charging to use it.

The launch of Qwen4-Max is probably the biggest AI news today, but other Chinese companies aren’t sitting idle. Tencent’s announcement isn’t quite on the same scale as Alibaba’s new flagship, although Hy4 Flash has some interesting numbers of its own.

The details: Hy4 Flash uses a Mixture-of-Experts (MoE) architecture with just 12 billion active parameters. That’s considerably less than the full Hy4, which activates 49 billion out of 770 billion parameters.

Tencent has also made several changes to reduce the resources needed to run the model:

  • Memory: Its KV cache is six times smaller than Hy4’s.
  • Speed: Tencent claims generation speeds exceeding 150 tokens per second.
  • Efficiency: Post-training improvements reportedly reduced output token consumption by around 50%.

By the numbers: Tencent tested Hy4 Flash against several competing models using 85 internal WorkBuddy users across 259 tasks. Participants rated the results on a four-point satisfaction scale.

  • DeepSeek V4.1 Flash: 3.09, costing 1.20 yuan per task.
  • Claude Haiku 5.5: 2.81, costing 4.73 yuan.
  • Hy4 Flash: 2.79, costing just 0.26 yuan.
  • GLM-5.3 Flash: 2.53, costing 0.81 yuan.

Hy4 Flash came remarkably close to Claude Haiku 5.5 in this evaluation, while costing around 18 times less per task.

Yes, but: These are Tencent’s own evaluations. And the smaller model isn’t nearly as capable as the full Hy4 on some coding benchmarks. On Terminal-Bench 4.0, it scored 19.2 against Hy4’s 44.4.

Pricing and availability: Hy4 Flash is available through Tencent Cloud TokenHub at $0.07 per million input tokens and $0.281 per million output tokens. Cached input costs just $0.003 per million tokens.

Tencent is offering a 50% API discount until November 30, alongside free access through WorkBuddy. It also plans to release the model’s weights, although there’s no confirmed date yet.

Efficienist Newsletter

Get the core business tech news delivered straight to your inbox. We track AI, automation, SaaS, and cybersecurity so you don't have to.

Just read what you want, and be done with it.

Read Next