Reassessing Tencent AI Value: '40% Efficiency Boost + Extreme Pricing', Hunyuan HY3.0 Embraces 'Pragmatism' and Already Possesses Scalable Commercial Capability

TENCENT's latest released preview version of Hunyuan HY3.0 is shifting its AI strategy from a technological race to commercial implementation.

According to ZhuiFeng Trading Desk, a Citigroup research report indicates that TENCENT officially launched the Hunyuan HY3.0 preview version on April 23, 2026, on mainstream developer platforms such as GitHub and HuggingFace. The model adopts a Mixture of Experts (MoE) architecture, boasts 295 billion parameters, achieves a 40% improvement in inference efficiency compared to the previous generation, and enters the enterprise market with highly competitive pricing—with input costs as low as 1.2 RMB per million tokens and output costs at 4 RMB per million tokens.

Although named the "preview version," this model has already been immediately integrated into multiple core products under TENCENT, indicating it possesses the capability to solve real-world problems. This model marks a new stage where TENCENT's AI capabilities have entered the phase of scalable commercialization.

Citigroup maintains a Buy rating on TENCENT with a target price of HKD 783, corresponding to an expected total return of approximately 59.3% based on the current share price (HKD 495.20). HY3.0's "pragmatic" route—focusing on balancing quality, speed, and cost rather than simply pursuing laboratory scores—is the correct strategic direction. It is expected that as the official release of HY3.0 or subsequent iterations (such as 3.1/3.2) are launched, the model's capabilities will further improve.

Technical Architecture: MoE Model Balances Performance and Efficiency

HY3.0 is an MoE language model integrating fast and slow thinking, with a total parameter count of 295 billion, activated parameters of 21 billion, and support for an ultra-long context window of up to 256k. TENCENT has comprehensively rebuilt the underlying infrastructure for this model, from pre-training to reinforcement learning.

In terms of capabilities, HY3.0 has achieved significant breakthroughs in three core directions. First, in complex context processing, TENCENT proposed two evaluation benchmarks, CL-bench and CL-bench-Life, specifically designed to assess the model's context learning and instruction-following abilities. Second, in complex reasoning, HY3.0 performed exceptionally well in high-difficulty science and engineering reasoning tasks such as FrontierScience-Olympiad, IMOAnswerBench, Tsinghua University Mathematics PhD Entrance Exam, and the National High School Biology Olympiad. Third, in code and agents, HY3.0 achieved competitive results in benchmark tests including SWE-Bench Verified, Terminal-Bench 2.0, BrowseComp, and WideSearch.

The model has made its weights available on platforms such as GitHub, HuggingFace, ModelScope, and GitCode, and supports mainstream inference frameworks like vLLM and SGLang, further lowering the deployment threshold for developers.

Commercial Implementation: Immediate Integration of Over Ten Core TENCENT Products

The strategic significance of the HY3.0 preview version lies not only in its technical parameters but also in the depth of its immediate integration with TENCENT's existing product ecosystem.

According to the report, the HY3.0 preview version has been integrated into products including Yuanbao, ima, CodeBuddy, WorkBuddy, QQ, QQ Browser, Tencent Docs, Tencent Maps, Tencent Lexiang, and Tencent eSign, and is currently being promoted to more products such as WeChat Official Accounts, Peacekeeper Elite, and Tencent News.

This extensive product integration serves as a key indicator that the model possesses real-world application capabilities. Compared to model releases that remain solely on developer platforms, HY3.0's implementation path directly reaches TENCENT's massive user base, providing a clear channel for the commercial monetization of AI capabilities.

Pricing Strategy: Extreme Cost-Effectiveness Targets the Enterprise Market

On the cost and pricing front, the HY3.0 preview version achieved a 40% improvement in inference efficiency compared to previous models and offers competitive pricing schemes on Tencent Cloud Token Hub.

Under the pay-as-you-go model, the input price is as low as 1.2 RMB per million tokens, and the output price is 4 RMB per million tokens. For high-frequency users, Tencent Cloud also launched the TokenPlan package, with the personal version starting at 28 RMB per month.

HY3.0 clearly positions itself to help enterprises deploy AI applications in intelligent agent workflow environments where costs are controllable and operations are stable, rather than simply pursuing extreme scores in laboratory evaluations. This "pragmatic" positioning aligns closely with the core demands of current enterprise customers regarding AI implementation, helping TENCENT expand its market share in the cloud services and AI commercialization sectors.