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。
Arcee[2] 的文章《Trinity Large: An Open 400B Sparse MoE Model》日期为2026年1月27日,其中写道“We're putting out three variants”;Hugging Face[4] 上的 Trinity-Large-TrueBase、Base 和 Preview 仓库均在当天创建。
Arcee's[1] April follow-up: Preview was a light instruct post-train, crossed 3.37 trillion tokens served on OpenRouter in two months, and was succeeded by the reasoning Trinity-Large-Thinking.
The technical report submitted 19 February 2026: 400B total and 13B active parameters, 17T tokens, SMEBU expert load balancing, Muon optimizer, no loss spikes, plus the Trinity Mini (26B, 3B active) and Nano (6B, 1B active) models.[1]
Model card: about 398B total and 13B active, 256 experts with 4 active, 6 dense layers, 17T pretraining tokens on 2,048 B300 GPUs, 20B tokens of instruction tuning, and MMLU 87.2 and AIME 2025 24.0 against Llama 4 Maverick's 85.5 and 19.3.[1]
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: Trinity Large Preview, Base and TrueBase Released