Tencent Hy3 Targets AI Agents Over Bigger Language Models

Tencent has officially introduced Hy3, the third generation of its flagship large language model, signaling a strategic shift toward practical AI applications instead of competing solely on model size. Unveiled last week, the new model is designed to power AI agents, coding assistants and enterprise productivity tools, reflecting a broader trend among Chinese technology companies to prioritize deployment efficiency and commercialization as hardware limitations continue to influence AI development.

Hy3 is built on a Mixture-of-Experts (MoE) architecture featuring 295 billion total parameters, with 21 billion active parameters and a 256K context window. Rather than focusing on benchmark dominance, Tencent says the model is optimized for real-world enterprise workflows and AI-powered automation.

Independent testing indicates Hy3 performs strongly in agent-focused tasks. According to AI consultancy Flowtivity, the model achieved a score of 84.2 on BrowseComp and 79.1 on the public MCP-Atlas benchmark, placing it alongside leading proprietary models such as Claude Opus 4.8 and GPT-5.5 for agentic search and tool orchestration. Hy3 also recorded a 5.4% hallucination rate, significantly lower than Grok 4.5’s 54% and competitive with other frontier AI systems.

Its performance is less dominant in advanced coding evaluations. Hy3 scored 78% on SWE-bench Verified, behind GLM-5.2’s 84.2% as well as Claude Opus 4.8 and GPT-5.5. The gap becomes more noticeable on demanding benchmarks, including Terminal-Bench 2.1, where Hy3 scored 71.7 compared with GLM-5.2’s 81, and DeepSWE, where it posted 28.0 against GLM-5.2’s 46.2. Analysts attribute much of this difference to architecture, as GLM-5.2 uses a 744-billion-parameter MoE model with approximately 40 billion active parameters, nearly double Hy3’s active compute. One independent analysis concluded, “For the model size—only 21B active parameters—the results are remarkable.”

Tencent said Hy3 was developed using a “Co-Design” philosophy, in which AI models evolve alongside AI-native applications. Products such as WorkBuddy, Yuanbao, ima, Marvis and CodeBuddy serve as real-world testing environments, continuously supplying workflow data to improve the model. The company reported that WorkBuddy’s internal task success rate increased from 72% to 90%, while average execution time dropped by 34%. Tencent also said hallucination rates in Yuanbao’s long-document processing and AI search features were reduced by more than half.

The company has priced Hy3 at approximately $0.18 per million input tokens and $0.59 per million output tokens through Tencent Cloud. Tencent also offers an FP8-quantized version that fits on a single 8x H200 node using less than 300GB of memory, making self-hosting a practical option for enterprises with data sovereignty requirements.

Hy3’s Mixture-of-Experts architecture activates only a portion of its parameters for each token, allowing between three and eight times greater GPU throughput than dense models in large-scale deployments, according to infrastructure analysis. However, industry experts note that MoE systems face challenges including expert underutilization and load-balancing complexity, while dense models can remain more predictable and economical for lighter enterprise workloads.

Tencent believes its extensive software ecosystem will overcome those limitations by continuously generating real-world usage data. Applications such as WorkBuddy, Yuanbao, WeChat, gaming services and productivity platforms provide diverse interaction patterns that refine the model over time. The company said daily Hy3 token usage has increased twentyfold since its preview release, while the number of users actively selecting Hy3 within WorkBuddy has grown sixfold.

Tencent’s strategy mirrors a broader shift in enterprise AI. Anthropic has reportedly surpassed OpenAI in enterprise API market share during 2026, holding roughly 32% compared with OpenAI’s 25%, largely by emphasizing coding reliability and long-context reasoning instead of model size. Anthropic’s Claude Code coding agent has also become a major revenue driver, reportedly reaching $2.5 billion in annualized revenue. Both companies are betting that enterprise customers place greater value on workflow completion, reliability and low latency than on marginal improvements in academic benchmarks.

Tencent’s latest release also highlights a broader transformation in enterprise software, where office productivity platforms are evolving into AI-powered execution engines. With applications such as WorkBuddy already capable of automating scripts and orchestrating workflows, Tencent believes its integrated ecosystem can continuously improve Hy3 using millions of real business tasks. The company sees this product-driven feedback loop as a competitive advantage that could help narrow the capability gap with leading Western AI models despite ongoing hardware constraints.