GPT-6 后中国模型沉寂一月,GLM-Omni 与 Kimi K3.1 曝光
Chinese models went quiet for over a month following the release of GPT-6 and Fable 5.5. Some believe China can no longer rely on distillation to close the gap
GPT-6 和 Fable 5.5 发布后中国模型沉寂一个多月,有人认为中国已无法靠知识蒸馏追平 OpenAI 与 Anthropic。唐杰参与合著的论文提到原生多模态模型 GLM-Omni(推测为 GLM-5.5),通过图文交错对齐将任务准确率从 12.57% 提升至 26.86%。SGLang 提交中出现 Kimi K3.1 引用,保留 K3 tokenizer 与对话格式并新增“medium”推理档位,外界关注预计 Q4 的 Qwen 4 和预计 2027 年的 Hunyuan 4.1。
正文
当前显示原文,中文全文译文尚未就绪。
Chinese models went quiet for over a month following the release of GPT-6 and Fable 5.5. Some believe China can no longer rely on distillation to close the gap with OpenAI and Anthropic. They are likely wrong.
A recent paper co-authored by https://t.co/0DZXxqJWrG founder Tang Jie mentions GLM-Omni (speculated to be GLM-5.5), a native multimodal model capable of understanding and generating images. Through interleaved image–text alignment, its task accuracy reportedly jumped from 12.57% to 26.86%.
Meanwhile, Kimi K3.1 references have surfaced in SGLang commits. It retains the K3 tokenizer and chat format while introducing a new "medium" reasoning tier. All eyes are now on Qwen 4 (expected in Q4) and Hunyuan 4.1 (expected in 2027).