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[5] 的公告 'Mistral Small 3' 日期为 2025年1月30日,其更新日志写明 'January 30: We released Mistral Small 3 (mistral-small-2501)';Simon Willison[3] 的文章日期相同。
Restates Mistral's[1][5] claims: aimed at about 80% of generative AI tasks, fewer layers than competitors for lower latency, matches Llama 3.3 70B instruct at more than three times the speed, and was not trained with RL or synthetic data, so it can seed reasoning models.
Same-day write-up: the license move from the Mistral[1][5] Research License to Apache 2.0, the roughly 20GB class of model that suits laptops, the API price of $0.10 per million input tokens and $0.30 output (half the old Small price, against $0.15 and $0.60 for GPT-4o mini), and that it is the fourth Mistral Small release.
Mistral's[1][5] card: 24B parameters, 32k context, 131k-vocabulary Tekken tokenizer, Apache 2.0, about 55 GB of GPU RAM in bf16, and internal scores of MMLU-Pro 0.663, GPQA 0.453, HumanEval 0.848, MATH 0.706, IFEval 0.829, MT-Bench 8.35, WildBench 52.27 and Arena Hard 0.873.
Mistral's[1] changelog entry 'January 30: We released Mistral Small 3 (mistral-small-2501)'.
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 Small 3 (24B) Released