Hy3 preview is a 295B-parameter Mixture-of-Experts (MoE) language model from Tencent Hunyuan, built for production-grade agent workloads. With only 21B parameters activated per token and native 256K context support, it handles complex tasks like cross-file code refactoring, long-document analysis, and multi-step tool use, rather than just generating fluent dialogue. Hy3 scores near state-of-the-art on SWE-bench Verified and advanced STEM benchmarks, while offering three inference modes (no_think, think_low, think_high) to dynamically trade off latency and reasoning depth. Its sparse activation architecture delivers competitive intelligence at a significantly lower token cost....

Hunyuan-A13B-Instruct 僅啟用其 80 B 參數中的 13 B，卻能在主流基準上匹敵更大的 LLMs。它提供混合推理：每次呼叫可切換為低延遲“快速”模式或高精度“慢速”模式。內建 256 K-token 上下文，允許它在不減低功效的情況下解析書籍長度的文件。代理技能為 BFCL-v3、τ-Bench 和 C3-Bench 領導力而調校，使其成為優秀的自主助手基礎。分組查詢注意力和多格式量化提供記憶體輕量、GPU 高效的推理，適合現實世界的部署，並具備內建多語言支持和堅固的安全對齊，適用於企業級應用。...