Tencent launched Hunyuan Hy3, a 295B MoE model with 21B active parameters, claiming 90% agent task resolution, surpassing DeepSeek V4 Pro and Qwen 3.7 Max.
Tencent launched Hunyuan Hy3, a 295B MoE model with 21B active parameters, claiming 90% agent task resolution. The model surpasses DeepSeek V4 Pro and Qwen 3.7 Max on key benchmarks, per Tencent's internal evaluations.
Key facts
- 295B total parameters, 21B active parameters
- 90% agent task resolution claimed
- 14:1 sparsity ratio for inference efficiency
- Surpasses DeepSeek V4 Pro and Qwen 3.7 Max
- Launched amid Chinese agent regulatory crackdown
Tencent's Hunyuan Hy3, a 295B MoE model with 21B active parameters, achieves 90% agent task resolution, according to Tencent's official launch. The model surpasses DeepSeek V4 Pro and Qwen 3.7 Max on key benchmarks, though Tencent did not disclose independent third-party verification of the 90% claim. The 21B active parameters represent a 14:1 sparsity ratio, a pragmatic choice for inference cost at scale.
Timing: Agent Regulatory Crackdown
The launch comes as ByteDance's Doubao and Alibaba's Qwen sunset agent features due to regulatory compliance, per SCMP. Doubao's agent feature goes offline July 15, with Qwen following by October 15, as Beijing enforces new rules on humanlike AI interaction services. Tencent's Hy3 launch appears to bet on enterprise agent use cases where regulatory risk is lower than consumer-facing agents.
Benchmark Context
Hy3's claimed 90% agent task resolution compares favorably to DeepSeek V4 Pro and Qwen 3.7 Max, though benchmark methodology matters. Agent task resolution typically measures success rate on multi-step tool-use tasks like web navigation and API calls. DeepSeek V4 Pro, launched in June 2026, scored 85% on similar internal evaluations. Qwen 3.7 Max, Alibaba's flagship, scored 82%. The 5-8 point gap is significant but unverified by third parties.
Architecture Details
The 295B total parameter count with 21B active parameters follows the MoE trend set by DeepSeek-V3 (671B total, 37B active) and Qwen 3.6 Plus (285B total, 24B active). Tencent's 14:1 sparsity ratio is more aggressive than DeepSeek's 18:1, suggesting a focus on inference efficiency over raw parameter count. The model likely uses a top-2 routing mechanism, standard for MoE architectures, though Tencent did not confirm routing details.
Monetization Strategy
Tencent positions Hy3 for enterprise agent use cases, including customer service, code generation, and data analysis. The model is available via Tencent Cloud's API, with pricing undisclosed. This contrasts with the consumer agent retreat by ByteDance and Alibaba, suggesting Tencent sees enterprise agents as a more viable monetization path. The Chinese enterprise AI market is projected to reach $15B by 2027, per IDC data cited in the source.
Competitive Landscape
Hy3 enters a crowded field: DeepSeek V4 Pro (June 2026), Qwen 3.7 Max (May 2026), and ByteDance's Doubao Pro (April 2026). DeepSeek recently raised $7B in its first major funding round, abandoning its no-funding pledge, while Alibaba blocked Claude Code after Anthropic's tracking experiment. Tencent's model launch signals it will compete on agent performance rather than price, given its massive WeChat ecosystem for distribution.
What to watch
Watch for third-party benchmark results from SWE-Bench or GAIA on Hy3's agent performance, expected within 30 days. Also monitor Tencent Cloud's agent API pricing and whether enterprise adoption crosses 10K active deployments by Q4 2026.
Source: pandaily.com
Originally published on gentic.news
Top comments (0)