Tencent release Hy3

Tencent’s AI division, the Hy team, has officially released Hy3 (Hunyuan Hy3), its final, production-grade foundation model. Upgraded significantly from the preview model launched in April, Hy3 focuses heavily on real-world utility, autonomous AI agents, and production-grade reliability.

1. The Architecture: Massive Yet Lightweight

Hy3 uses a highly efficient, sparse Mixture-of-Experts (MoE) framework:

• Total Parameters: 295 Billion.
• Active Parameters: 21 Billion parameters per token. By firing only a fraction of its total network (top-8 routing across 192 experts), it delivers flagship-level performance at a fraction of the compute cost.
• Context Window: 256K tokens, allowing the model to parse massive codebases, multi-turn corporate chats, or entire research documents without dropping context.
• Multi-Token Prediction (MTP): Includes a 3.8B parameter MTP layer that predicts multiple tokens simultaneously, dramatically boosting generation speeds via speculative decoding in vLLM and SGLang.

2. Key Capabilities & Performance

Tencent took a pragmatic approach with Hy3, optimizing it for system stability and reducing hallucinations over simply chasing raw benchmark scores:

• 90% Agent Task Resolution: In internal tests on Tencent’s enterprise platform, WorkBuddy, Hy3 achieved a 90% task completion rate—a massive leap from the 72% managed by the preview version—while cutting down processing time by 34%.
• Halving Hallucination Rates: Through heavy post-training and reinforcement learning (RL), Tencent slashed the model’s structural hallucination rate down by half to 5.4%.
• Benchmark Surpasses: On reasoning and coding benchmarks, Hy3 scored 68.5 on the ClawEval metric, pushing past competitive models like DeepSeek V4 Pro and Qwen 3.7 Max. Its web-browsing capability (BrowseComp at 84.2) closely rivaled proprietary flagships like GPT-5.5.

3. Deployment and Open-Source Strategy

Tencent is aiming for immediate developer adoption with highly accessible licensing and pricing structures:

• Open-Source & Hugging Face: Released under the Apache License 2.0 for full commercial availability. The full FP16 weights sit at 598GB on Hugging Face, alongside a highly optimized 300GB FP8 quantized checkpoint for cheaper corporate hosting.
• Consumer Integration: Tencent immediately plugged Hy3 into its consumer-facing AI assistant, Yuanbao, powering an all-new free AI-agent creation feature for users.
• Aggressive Cloud Pricing: For external developers using Tencent Cloud’s APIs, the pricing has been aggressively set at 1 yuan per million input tokens, 4 yuan per million output tokens, and an incredibly cheap 0.25 yuan per million tokens for cached context.