Tencent's Hy3 Model Prioritizes AI Agents Over Parameter Count

A Different AI Race

Tencent has released Hy3, a large language model that reflects a strategic divergence from Silicon Valley's obsession with parameter count. The third-generation model features a Mixture-of-Experts architecture with 295 billion total parameters and 21 billion active parameters, but its positioning reveals more than its specifications: Tencent is optimizing for real-world AI agents and enterprise productivity rather than benchmark dominance.

The release, first reported by Forbes contributor Vivian Toh, comes as Chinese AI companies increasingly prioritize deployment efficiency and commercialization over raw model size—a shift driven partly by ongoing hardware constraints in the domestic market.

Why it matters

Tencent's approach tests whether product-integrated AI development can compensate for capability gaps with frontier Western models. With millions of daily users across WeChat, WorkBuddy, and other applications generating training data, the company has built a feedback loop that standalone model providers cannot easily replicate. If successful, this strategy could establish a viable alternative to the compute-intensive scaling paradigm dominating Western AI development.

Performance Trade-offs

Independent evaluations from AI consultancy Flowtivity show Hy3 excels at agentic tasks. The model scored 84.2 on BrowseComp and 79.1 on MCP-Atlas, competing with Claude Opus 4.8 and GPT-5.5. Its hallucination rate of 5.4% significantly undercuts Grok 4.5's 54% and matches leading proprietary models.

Coding benchmarks reveal limitations. Hy3 achieved 78% on SWE-bench Verified, trailing GLM-5.2's 84.2% and frontier models. The gap widens on demanding tests: 71.7 versus GLM-5.2's 81 on Terminal-Bench 2.1, and 28.0 versus 46.2 on DeepSWE. GLM-5.2's 744-billion-parameter architecture with roughly 40 billion active parameters—nearly double Hy3's activated compute per token—explains much of the difference.

The Ecosystem Advantage

Tencent describes Hy3's development as "Co-Design," where models and applications evolve together. Products including WorkBuddy, Yuanbao, and CodeBuddy serve as live testing environments. The company reports WorkBuddy's internal task success rate increased from 72% to 90%, while execution time dropped 34%. Yuanbao's hallucination rates in long-document scenarios fell by more than half.

Daily token consumption of Hy3 has increased twenty-fold since preview release, according to Tencent, while users actively selecting Hy3 inside WorkBuddy grew six times. This usage generates diverse interaction patterns that continuously refine the model.

Pricing reflects the efficiency focus: approximately $0.18 per million input tokens and $0.59 per million output tokens via Tencent Cloud. An FP8-quantized variant fits on a single 8x H200 node under 300GB, making self-hosting viable for enterprises.

Not Just a China Story

Tencent's strategy parallels developments in Western markets. Anthropic has captured roughly 32% of enterprise API market share versus OpenAI's 25% in 2026 by emphasizing coding reliability and long-context reasoning over scale. Claude Code, its terminal-native coding agent, reportedly reached $2.5 billion in annualized revenue.

Both companies are betting that enterprise customers prioritize workflow completion, reliability, and latency over marginal benchmark gains—suggesting the product-first approach transcends geographic boundaries.

Open Questions

Whether Tencent's data flywheel can close capability gaps with larger models remains uncertain. The early results show strength in agentic tasks and reliability, but the coding performance differential indicates limits to the efficiency-first approach. The answer will determine whether Hy3 represents a competitive alternative or simply a well-integrated domestic solution.

These details were first reported by Vivian Toh for Forbes.