| 规格 | 数值 |
|---|---|
| 发布方 | Google DeepMind |
| 模型 | Ultra、Pro 和 Nano |
来自 Google 发布文章及其文本基准测试表图片(GPT-4 的 MMLU 为表格脚注中的 CoT@32 数值)。文章未给出上下文窗口、价格或最大输出。 · 来源
Estimate from the version's published specs, benchmarks and reviews, 0-5; scores are only comparable within one criteria year. Scored as launched: Gemini 1.0 family, 6 Dec 2023; Pro in Bard then, API from 13 Dec, Ultra unreleased. Existing axes preserved: reasoning/coding/speed/price null, 32k context=0, family modalities=5 (Bard shipped text-only). Reviews concern Gemini Pro or launch analysis, not an independently tested Ultra. Round21 replaces conflicting-date Akkio analysis with Akter et al. arXiv v1 (18 Dec 2023), a transparent ten-dataset comparison with code. Gemini Pro is slightly below GPT-3.5 Turbo overall, with task-specific strengths. This small early API-era study does not establish current model performance or anchor-compatible scores.[1]
I’d say both models are on the same level.
The updated version of Bard did well with some travel tasks, particularly finding flight options, but its answers still need to be taken with a grain of salt.
I love that ‘above 90%’ turns out to be exactly 90.04%, whereas human expert is 89.8%, prior SOTA was 86.4%.
Gemini Pro achieves accuracy that is close but slightly inferior to the corresponding GPT 3.5 Turbo
| Gemini 1.0 · 2023年12月7日 | 0.60 |
|---|---|
| Google Bard’s Gemini Upgrade: We Tested it for Travel · 2023年12月9日 | 0.00 |
| A Hands-On Comparison: Gemini Pro vs GPT-3.5 · 2023年12月12日 | 0.20 |
| An In-depth Look at Gemini's Language Abilities · 2023年12月18日 | -0.25 |
这个标记会跟进新闻:它显示目前已知的情况,这里列出改了什么以及起因。
读者上传的照片,与图钉带来源的媒体分开显示。你可以为吃到或看到的内容评分,在餐厅图钉上还能直接在照片上标记菜品。
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。
Google 于 2023 年 12 月 6 日发布的帖子“Gemini 简介:我们最大、能力最强的 AI 模型”(Pichai 和 Hassabis 在虚拟新闻发布会上宣布了这一消息); Gemini Pro 于 12 月 13 日到达巴德,Ultra 则于 2024 年 2 月 8 日到达巴德。
Opened source supports this specific review extract; publication semantics and any live-benchmark scope are recorded in the research catalog. Does not establish pin event date. Review verification 2026-10-27 (round 21): Excluded from native reviews. Akkio original publication chronology remains conflicting. Replace active citation with version-fixed original third-party Gemini Pro benchmark paper; this does not resolve the old article date.
The Wikipedia article on the Gemini family dates the 1.0 announcement to 6 December 2023 and lists Nano (6 Dec), Pro (13 Dec) and Ultra (8 Feb 2024) as separate availability dates.
Review evidence round21: Original arXiv v1 submission 2023-12-18 18:47:42 UTC, not Dec24 v2 revision. Abstract and linked code identify ten datasets and Gemini Pro; no Ultra test or present-day Gemini inferred. Passage sentiment is a modest negative comparative verdict, not a benchmark-derived axis. Review verification 2026-10-27 (round 21): Original arXiv v1 submission 2023-12-18 18:47:42 UTC, not Dec24 v2 revision. Abstract and linked code identify ten datasets and Gemini Pro; no Ultra test or present-day Gemini inferred. Passage sentiment is a modest negative comparative verdict, not a benchmark-derived axis.
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: Gemini 1.0 Released