Trending: On-device modelsSearch
iHeartGeek
iTECH

Tencent launches Hy4 flash with million-token context and lower API prices

Tencent says Hy4 flash combines a million-token context window with multimodal input, fast decoding and sharply lower API prices.

Tencent's official Hy4 flash API pricing comparison with Hy4

Tencent has launched Hy4 flash, a new addition to its Hy family of AI models that is designed to make long-context and multimodal workloads cheaper to run. The company says the model has 224 billion total parameters, including 64 billion Engram parameters, with 12 billion active parameters, and supports context windows of up to 1 million tokens.1

The headline pricing is US$0.070 per million input tokens, US$0.281 per million output tokens and US$0.003 per million cached input tokens.1 Tencent presents those figures as a substantial reduction from Hy4, with the biggest difference appearing in cached input and output costs. The pricing table in the company's announcement puts Hy4 flash well below Hy4 across all three categories.

Tencent says Hy4 flash uses hybrid attention, LatentMoE and Engram features during pre-training, while reinforcement learning reduces average end-to-end task output by 50 per cent. The model is also claimed to decode at more than 150 tokens per second through a combination of megakernel optimisation, 4-bit quantisation and DFly techniques.1

The company is positioning Hy4 flash for practical work rather than only benchmark tests. It says the model is available through Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, with OpenRouter integration planned. WorkBuddy access is free until 30 November, while Tencent Cloud TokenHub customers can receive 50 per cent off standard pricing during the same period.1

Tencent's performance comparisons need the usual caveat: its WorkIndex benchmark and blind evaluation were conducted using internal tasks and users, so they are company-reported rather than an independently reproduced result. The launch is still notable because it links a million-token context window to a pricing model aimed at everyday document, presentation and media-processing work.

Our opinion

Hy4 flash looks less like a flagship announcement and more like a pricing and deployment play. Tencent is trying to make a capable long-context model easier to use repeatedly, where cached prompts and high-volume workflows matter more than a single headline benchmark score. The important test will be whether developers find the model reliable enough to leave in production after the introductory discounts end. A cheap million-token window is useful; a predictable one is what turns it into infrastructure.