| 规格 | 数值 |
|---|---|
| 发布方 | Google DeepMind |
| 规模 | 1B、4B、12B 和 27B |
| 上下文窗口 | 128k tokens(1B 为 32K) |
| 输入 | 文本和图像(1B 仅文本) |
来自 Google 的 Gemma 3 模型卡和开发者指南;图表按模型卡中指令微调模型的结果比较四种规模。模型卡未给出 Gemma 2 数据、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 (12 Mar 2025) from Google's developer guide opened this run: 128k-token context for the 4B/12B/27B (32K on the 1B) = 1; vision-language input with text output on all sizes but the 1B = 3. Reasoning, coding, price and speed are null: the Artificial Analysis Intelligence Index page for Gemma 3 27B shows no index value in this run, Terminal-Bench is not published, Google sets no price for open weights and no launch-time speed is published. Three reviews found: two independent (Simon Willison's hands-on notes; TechTarget with Forrester and Futurum analysts) plus one vendor test (Novita AI, a hosting reseller that ran real prompts). Searches also covered general review queries and Artificial Analysis; no further independent hands-on write-ups were found.[5]
The one big miss is that it ignored my instructions to include illustrative quotes—I don’t know if modifying the prompt will fix that but it’s disappointing that it didn’t handle that well, given how important direct quotes are for building confidence in RAG-style responses.
The massive reduction in resource requirements is probably the biggest or most interesting component of this
It appears that the model did not fully adhere to prompt. Instead, it generated a spinning ring with alphabets.
| Notes on Google’s Gemma 3 · 2025年3月12日 | 0.10 |
|---|---|
| Google’s Gemma 3 can run on a single TPU or GPU | TechTarget · 2025年3月12日 | 0.60 |
| Gemma 3 27B on Novita AI: Really a Single-GPU Model? - Novita · 2025年3月31日 | -0.40 |
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。
谷歌的博文 'Introducing Gemma 3: The most capable model you can run on a single GPU or TPU' 和开发者博客指南的日期均为2025年3月12日,并称Gemma 3当天即可使用。
Router listing of the 4B instruct model: 131,072-token context, text and image input, $0.05 per 1 million input tokens and $0.10 per 1 million output tokens.[1]
Dates the family: Gemma February 2024, Gemma 2 on 27 June 2024, Gemma 3 on 12 March 2025, Gemma 4 on 2 April 2026; 128K context for every size except the 32K 1B.[1]
The technical report: higher local-to-global attention ratio to cut KV-cache memory, distillation, and the claim that Gemma3-4B-IT is competitive with Gemma2-27B-IT; Table 6 holds the instruction-tuned benchmark numbers.[1]
Developer guide: four sizes each as pre-trained and instruction-tuned, 128k context, 140+ languages, 2T/4T/12T/14T training tokens, SigLIP vision encoder, distillation plus RLHF, RLMF and RLEF post-training, LMArena 1338 for the 27B.[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: Gemma 3 4B, 12B, 27B and 1B Released
| 基准 | Gemma 2 2B | Gemma 2 9B | Gemma 2 27B | Gemma 3 1B | Gemma 3 4B | Gemma 3 12B | Gemma 3 27B |
|---|---|---|---|---|---|---|---|
| MMLU-Pro | 15.6 | 46.8 | 56.9 | 14.7 | 43.6 | 60.6 | 67.5 |
| LiveCodeBench | 1.2 | 10.8 | 20.4 | 1.9 | 12.6 | 24.6 | 29.7 |
| Bird-SQL (dev) | 12.2 | 33.8 | 46.7 | 6.4 | 36.3 | 47.9 | 54.4 |
| GPQA Diamond | 24.7 | 28.8 | 34.3 | 19.2 | 30.8 | 40.9 | 42.4 |
| SimpleQA | 2.8 | 5.3 | 9.2 | 2.2 | 4.0 | 6.3 | 10.0 |
| FACTS Grounding | 43.8 | 62.0 | 62.4 | 36.4 | 70.1 | 75.8 | 74.9 |
| Global MMLU-Lite | 41.9 | 64.8 | 68.6 | 34.2 | 54.5 | 69.5 | 75.1 |
| MATH | 27.2 | 49.4 | 55.6 | 48.0 | 75.6 | 83.8 | 89.0 |
| HiddenMath | 1.8 | 10.4 | 14.8 | 15.8 | 43.0 | 54.5 | 60.3 |
| MMMU (val) | - | - | - | - | 48.8 | 59.6 | 64.9 |