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。
Mistral[5] 的公告 'Introducing Mistral Large 4' 日期为 2026 年 10 月 6 日,描述的是公开预览;其文档中的模型卡将发布日期列为 2026 年 10 月 6 日(v26.10,公开预览),TechCrunch[2] 的报道也是同一天。
TechCrunch's 6 October report: about 1 trillion parameters, not yet open-weight and served only through a public guardrail endpoint, weights planned about three weeks after safety testing, trained on roughly 4,000 NVIDIA GPUs ('two to three times less than our Chinese competitors'), pitched at cybersecurity, finance and chip design, with ASML leading the Series C and Samsung the Series D.
Simon Willison's same-day note: 1 trillion total and 49 billion active parameters, preview on Mistral's[1][5] API with weights promised for the end of the month, only 'none' and 'high' reasoning levels, Artificial Analysis Intelligence Index 38 (just behind DeepSeek 4.1 Flash, against 9 for Mistral Large 3), and his view that Mistral is back to roughly six months behind the frontier.
Same-day report: 1 trillion parameters with 49 billion active, natively multimodal, a public preview API on Mistral[1][5] Studio running on the training infrastructure, 160+ languages, weights plus architecture details and post-training methodology due by the end of October, and the reinforcement-learning run behind the preview still in progress.
Mistral's[1] API model card: public preview v26.10 released October 6, 2026, an open-weight multimodal granular mixture of experts with 1.05T total and 52B active parameters, a 1.6B vision encoder, a 1M-token context and prices of $0.68 input and $2.09 output per million tokens (original $1.36 and $4.18).
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: Mistral Large 4 Released