EVENT DOSSIER
On August 29, 2026, Tencent launched the open-source text input large language model Hy4 Preview. This model has a total parameter size of 770B, an active parameter count of 49B, and 1M token context windows. Compared to Hy3 released in July (with a total parameter size of 295B), Hy4 has significantly increased in scale. Hy4 includes a chat_template.jinja provided by Hugging Face, which contains validation logic for the reasoning_effort parameter. It only supports two reasoning effort levels: ‘high’ and ‘no_think’, with ‘high’ being the default setting. Tests show that this model can generate images containing elements such as helmets when processing complex prompts (e.g., generating SVG images of seagulls riding a bike), along with truncated English reasoning trajectories.
腾讯今日发布新开源文本输入（无视觉）大语言模型 Hy4 Preview，拥有 770B 总参数量、49B 活跃参数量及 1M token 上下文窗口。该模型参数量较 7 月发布的 Hy3（295B 总参数量、21B 活跃参数量、256,000 上下文窗口）显著增加。Hy4 在 Hugging Face 提供 chat_template.jinja，包含对 reasoning_effort 参数的校验逻辑，仅支持"high"和"no_think"两个推理努力级别，其中"high"为默认设置。作者使用 OpenRouter 测试了"Generate an SVG of a pelican riding a bicycle"提示词，模型生成了包含头盔等元素的 SVG 图像，并附带了截断的英文推理轨迹。