Overview
Tencent Hy4 preview is an open-weight, preview-stage large language model designed for real-world productivity, including software development, office work, agent tasks and scientific research. Tencent released it on August 28, 2026. (tencent.com)
Capabilities and use
- Architecture: Mixture of Experts, with 770 billion total backbone parameters and 49 billion activated per token. Tencent also released an FP8 quantized variant. (github.com)
- Context: Tencent Cloud lists a 1M-token context window, with up to 960K input tokens and 64K output tokens. (tencentcloud.com)
- Features: The international TokenHub listing includes deep reasoning, structured output, function calling and caching. Applications include coding, document and data analysis, and research workflows. (tencentcloud.com)
- Access and license: Tencent provides model weights under Apache License 2.0. The announcement also identifies Tencent Cloud TokenHub and Tencent applications including WorkBuddy and CodeBuddy as access options. (github.com)
Limitations
Tencent describes this as an early version that can spend longer than necessary reasoning and may over-verify its work. The international model list does not list image understanding for the preview; Tencent says visual understanding was first introduced in the later Hy4 release. TokenHub International marks Hy4 preview as to be discontinued, without a date in the cited listing. Outputs should be verified, particularly for consequential tasks. (github.com)
API pricing
Prices below are per 1 million tokens for hy4-preview, checked November 16, 2026. Regional services and billing terms differ; these are not currency conversions of one plan.
- Tencent Cloud China, TokenHub online inference, Guangzhou: CNY 6 input, CNY 18 output, and CNY 0.3 cache-hit input. The separately listed batch inference rates are CNY 3 input, CNY 9 output, and CNY 0.15 cache-hit input. Official China pricing. (cloud.tencent.com)
- Tencent Cloud International, TokenHub online inference, Singapore: USD 0.834 input, USD 2.501 output, and USD 0.042 cache-hit input. Official international pricing. (tencentcloud.com)
Cache-hit rates apply only to qualifying cached input. Confirm current availability, regional rates and actual billing before use. (intl.cloud.tencent.com)