Tencent Open Sources Hy3 Preview 295B AI Model
Tencent Hunyuan said it has open-sourced Hy3 preview, a 295B A21B model. The preview was described as delivering improvements across chat, code, agents and math/science reasoning, with stronger instruction following and in-context learning, while boosting inference efficiency by 40%.
Hy3 preview is described as the first model to emerge from a rebuild of Tencent Hunyuan's pre-training and reinforcement-learning infrastructure that began in February 2026 and was designed around systematic capability, authentic evaluation and cost-effectiveness. The model is also described as fully integrated into Tencent's core businesses, with gains validated across multiple flagship AI products.
From the sources (3 posts)
@jeremyphowardRT @TencentHunyuan: 👋Hi /haɪ/, we're the Tencent Hy /haɪ/ team🐧 Today, we open source Hy3 preview (295B A21B), a leading reasoning and age…
@sairahul1This isn't random. Since February 2026, Tencent Hunyuan rebuilt their entire pre-training and RL infrastructure from scratch. Three principles they're building around: → Systematic capability (no one-trick ponies) → Authentic evaluation (
@swapnakpanda⬘ Hy3 preview delivers across-the-board improvements in chat, code, agents, and math/science reasoning. Outstanding instruction following and in-context learning. ⬙ Hy3 preview boosts inference efficiency by 40%, achieving the most competi