← All models

Hy3

Published Publisher Tencenthy3

Overview

Hy3 is Tencent’s production language model for practical reasoning and agent workflows, released on July 6, 2026. It succeeds Hy3 preview. Tencent describes it as a hybrid fast-and-slow-thinking mixture-of-experts model with 295 billion total parameters, 21 billion active parameters and a maximum context length of 256K tokens. (tencent.com)

Capabilities and uses

Hy3 supports reasoning, instruction following, in-context learning, code generation and agent tasks. Tencent reports improvements over the preview in performance, stability and cost efficiency. It is integrated into WorkBuddy, CodeBuddy, Yuanbao, Marvis and ima for uses including software development, workflow orchestration, knowledge-base question answering, document generation, file management and tool coordination. Some of these outcomes depend on the surrounding application and tools, not the model alone. (tencent.com)

Availability and limitations

Hy3 is available through Tencent Cloud TokenHub and as open weights under the Apache 2.0 license. Tencent Cloud describes a 256K-token context window, with TokenHub limits of 192K input tokens and 128K output tokens; these are separate platform limits within the overall context constraint. Self-hosting requires substantial compute: Tencent Cloud describes an eight-GPU recommended serving configuration. Outputs can still contain hallucinations or errors and should be checked, particularly before consequential use. (developer.cloud.tencent.com)

Official API pricing

Prices below are for the production hy3 model, not Hy3 preview. They are list prices per million tokens, checked on November 16, 2026. Domestic and international offerings are separate; the amounts are not currency conversions.

  • Tencent Cloud China, online inference, postpaid: CNY 1 input, CNY 4 output and CNY 0.25 for cache-hit input. The published tables list Guangzhou and Singapore regions. Official domestic pricing. (cloud.tencent.com)
  • Tencent Cloud China, batch inference: CNY 0.5 input, CNY 2 output and CNY 0.125 for cache-hit input per million tokens. The published tables list Guangzhou and Singapore regions. Official domestic batch pricing. (cloud.tencent.com)
  • Tencent Cloud international, online inference: USD 0.132 input, USD 0.528 output and USD 0.033 for cache-hit input per million tokens. The international pricing page lists regional offerings; check the selected region and account at purchase. Official international pricing. (intl.cloud.tencent.com)

Official release source

Tencent’s Hy3 announcement, July 6, 2026. (tencent.com)