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。
Mistral[6]的公告《Large Enough》日期为2024年7月24日,其更新日志写道“7月24日:我们发布了Mistral Large 2(mistral-large-2407)”;TechCrunch[4]在同一天发表了报道。
Lists Mistral[1][6] Large 2 as announced on July 24, 2024 with 123 billion parameters, a 128,000-token context and open weights under the Mistral Research License.
The Register's 25 July report: about 246GB of memory at 16-bit precision, deployable on one server with four or eight GPUs, weights on Hugging Face under the research license, and a 84% MMLU for Mistral[1][6] Large 2 against 88.6% for Llama 3.1 405B, 88.7% for GPT-4o and 88.3% for Claude 3.5 Sonnet.
TechCrunch's same-day report: Mistral's[1][6] new flagship, released a day after Llama 3.1 405B, with 123 billion parameters (under a third of the 405B), a 128,000-token context, 12 natural languages and 80 coding languages, no multimodal input, and a license under which any commercial application needs a paid license; Mistral had raised $640 million at a $6 billion valuation in June 2024.
Notes the 123B size, the 128,000-token context, that the research license is less open than Llama 3.1, that Mistral's[1][6] own benchmarks beat the larger Llama 3.1 405B on code and math, and a 69GB 4-bit build on Ollama.
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 2 Released