这些是读者照片,与图钉带来源的媒体分开显示。
第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均;资料每比最新的一份旧 180 天,权重减半。
Moonshot的技术博客以“Today, we are introducing Kimi K2.5”开篇,图片文件带有20260127的时间戳;Hugging Face[2]模型卡于2026-01-27公开,OpenRouter[3]当天上架kimi-k2.5,The Decoder[6]于1月28日报道。
Moonshot's model card: native multimodal agentic model, 1T total / 32B activated parameters, 256K context, MoonViT vision encoder, Modified MIT license and the K2.5 (Thinking) benchmark table against GPT-5.2 (xhigh), Claude 4.5 Opus, Gemini 3 Pro and DeepSeek V3.2.[1]
OpenRouter's record: created 2026-01-27, 262,144-token context, text and image input and third-party provider prices of $0.49 to $0.60 per million input tokens and $2.50 to $3.33 output.
17 February analysis of the PARL training method: only the orchestrator is trained while sub-agents stay frozen, the reward counters "serial collapse", and K2.5 beat GPT-5.2 Pro on BrowseComp and Claude Opus 4.5 on WideSearch.[1]
Kimi[1] Team's technical report (submitted 2 February 2026): joint text-vision pre-training, vision-free SFT and joint RL, plus Agent Swarm, which cuts latency by up to 4.5x against single-agent baselines.
有遗漏或错误吗?用你自己的话说明:能佐证此图钉的链接、不同的开始或结束日期及理由,或缺失、有误的信息。AI 会对照此图钉的来源进行核实,搜索更好的来源,并添加任何支持你说法的页面。图钉自身的来源仍然最重要。AI 也会查看图片:显示的是别的东西或显示效果差的图片会被移到后面或替换。
原标题: Kimi K2.5 Released