文件工作區

📝 編輯文件

把儲存操作固定在上方，避免被頁尾或編輯器工具列遮住。

瀏覽預覽返回列表刪除檔案

內容編輯區

Gemma 4 系列模型差異（概念整理）

Document

<p data-start="0" data-end="159"><span class="hover:entity-accent entity-underline inline cursor-pointer align-baseline"><span class="whitespace-normal">Google</span></span> 的 <strong data-start="40" data-end="51">Gemma 4</strong> 系列（延續自 Gemma 1 / 2）主要是「開放權重（open-weight）」的小到中型模型，設計重點是：<strong data-start="107" data-end="131">可本地運行 + 高效能推理 + 商用友好</strong>。不同版本之間的差異，主要在於「模型大小、能力、用途」。</p>
<p data-start="161" data-end="183">下面幫你整理重點差異（用最實用的角度看）👇</p>
<hr data-start="185" data-end="188">
<h1 data-section-id="1if1uh5" data-start="190" data-end="215">🔹 Gemma 4 系列模型差異（概念整理）</h1>
<h2 data-section-id="nbwavo" data-start="217" data-end="235">1️⃣ 模型尺寸（最核心差異）</h2>
<p data-start="236" data-end="255">通常會有幾種等級（實際命名可能類似）：</p>
<ul data-start="257" data-end="323">
<li data-section-id="ghe6dn" data-start="257" data-end="278"><strong data-start="259" data-end="278">Gemma 4 2B（小模型）</strong></li>
<li data-section-id="yvlxbg" data-start="279" data-end="300"><strong data-start="281" data-end="300">Gemma 4 7B（中模型）</strong></li>
<li data-section-id="vu5dv8" data-start="301" data-end="323"><strong data-start="303" data-end="323">Gemma 4 27B（大模型）</strong></li>
</ul>
<p data-start="325" data-end="331">👉 差異：</p>
<div class="TyagGW_tableContainer">
<div class="group TyagGW_tableWrapper flex flex-col-reverse w-fit" tabindex="-1">
<table class="w-fit min-w-(--thread-content-width)" data-start="333" data-end="487">
<thead data-start="333" data-end="351">
<tr data-start="333" data-end="351">
<th class="" data-start="333" data-end="338" data-col-size="sm">模型</th>
<th class="" data-start="338" data-end="343" data-col-size="sm">特點</th>
<th class="" data-start="343" data-end="351" data-col-size="sm">適合用途</th>
</tr>
</thead>
<tbody data-start="379" data-end="487">
<tr data-start="379" data-end="424">
<td data-start="379" data-end="384" data-col-size="sm">2B</td>
<td data-col-size="sm" data-start="384" data-end="407">超輕量、速度快、可在低階GPU/CPU跑</td>
<td data-col-size="sm" data-start="407" data-end="424">嵌入式、簡單聊天、工具輔助</td>
</tr>
<tr data-start="425" data-end="456">
<td data-start="425" data-end="430" data-col-size="sm">7B</td>
<td data-col-size="sm" data-start="430" data-end="441">平衡型（最常用）</td>
<td data-col-size="sm" data-start="441" data-end="456">本地AI助手、開發應用</td>
</tr>
<tr data-start="457" data-end="487">
<td data-start="457" data-end="463" data-col-size="sm">27B</td>
<td data-col-size="sm" data-start="463" data-end="471">高品質輸出</td>
<td data-col-size="sm" data-start="471" data-end="487">複雜推理、寫程式、長文本</td>
</tr>
</tbody>
</table>
</div>
</div>
<hr data-start="489" data-end="492">
<h2 data-section-id="15j2v9k" data-start="494" data-end="513">2️⃣ 指令版 vs Base版</h2>
<p data-start="514" data-end="529">Gemma 4 一定會分兩種：</p>
<h3 data-section-id="1pdw18a" data-start="531" data-end="547">▶ Base（基礎模型）</h3>
<ul data-start="548" data-end="614">
<li data-section-id="ujngl3" data-start="548" data-end="564">未對齊（沒有特別訓練成聊天）</li>
<li data-section-id="l9c825" data-start="565" data-end="614">適合：
<ul data-start="573" data-end="614">
<li data-section-id="1c75f74" data-start="573" data-end="587">自己 fine-tune</li>
<li data-section-id="1bym45q" data-start="590" data-end="614">做研究 / embedding / 特殊任務</li>
</ul>
</li>
</ul>
<h3 data-section-id="agr4bs" data-start="616" data-end="636">▶ Instruct（指令模型）</h3>
<ul data-start="637" data-end="687">
<li data-section-id="ll2338" data-start="637" data-end="657">已經做過對齊（類似 ChatGPT）</li>
<li data-section-id="6yxaeq" data-start="658" data-end="687">適合：
<ul data-start="666" data-end="687">
<li data-section-id="yia80b" data-start="666" data-end="670">聊天</li>
<li data-section-id="1tprjb" data-start="673" data-end="678">寫程式</li>
<li data-section-id="1oqbrdu" data-start="681" data-end="687">日常使用</li>
</ul>
</li>
</ul>
<p data-start="689" data-end="715">👉 一般人：<strong data-start="696" data-end="715">直接用 Instruct 就好</strong></p>
<hr data-start="717" data-end="720">
<h2 data-section-id="dyzg6m" data-start="722" data-end="742">3️⃣ 多模態能力（是否支援圖片）</h2>
<p data-start="743" data-end="759">Gemma 4 很可能延續趨勢：</p>
<ul data-start="761" data-end="803">
<li data-section-id="1ik1ctp" data-start="761" data-end="780">❌ 純文字版（Text-only）</li>
<li data-section-id="1uqxupx" data-start="781" data-end="803">✅ 多模態版（Text + Image）</li>
</ul>
<p data-start="805" data-end="811">👉 差異：</p>
<div class="TyagGW_tableContainer">
<div class="group TyagGW_tableWrapper flex flex-col-reverse w-fit" tabindex="-1">
<table class="w-fit min-w-(--thread-content-width)" data-start="813" data-end="891">
<thead data-start="813" data-end="824">
<tr data-start="813" data-end="824">
<th class="" data-start="813" data-end="818" data-col-size="sm">類型</th>
<th class="" data-start="818" data-end="824" data-col-size="sm">能力</th>
</tr>
</thead>
<tbody data-start="841" data-end="891">
<tr data-start="841" data-end="863">
<td data-start="841" data-end="853" data-col-size="sm">Text-only</td>
<td data-col-size="sm" data-start="853" data-end="863">專注文字處理</td>
</tr>
<tr data-start="864" data-end="891">
<td data-start="864" data-end="877" data-col-size="sm">Multimodal</td>
<td data-col-size="sm" data-start="877" data-end="891">可讀圖片、做視覺理解</td>
</tr>
</tbody>
</table>
</div>
</div>
<hr data-start="893" data-end="896">
<h2 data-section-id="1pr2gu0" data-start="898" data-end="926">4️⃣ 上下文長度（Context Window）</h2>
<p data-start="927" data-end="935">不同版本會有差：</p>
<ul data-start="937" data-end="975">
<li data-section-id="1i2mz5k" data-start="937" data-end="956">小模型：8K～32K tokens</li>
<li data-section-id="b6t1xx" data-start="957" data-end="975">大模型：可到 128K 甚至更高</li>
</ul>
<p data-start="977" data-end="983">👉 影響：</p>
<ul data-start="984" data-end="1012">
<li data-section-id="h4r5jk" data-start="984" data-end="994">能不能處理長文件</li>
<li data-section-id="yrivvy" data-start="995" data-end="1012">能不能做 RAG / 文件問答</li>
</ul>
<hr data-start="1014" data-end="1017">
<h2 data-section-id="11r81bt" data-start="1019" data-end="1036">5️⃣ 性能 vs 硬體需求</h2>
<div class="TyagGW_tableContainer">
<div class="group TyagGW_tableWrapper flex flex-col-reverse w-fit" tabindex="-1">
<table class="w-fit min-w-(--thread-content-width)" data-start="1038" data-end="1165">
<thead data-start="1038" data-end="1064">
<tr data-start="1038" data-end="1064">
<th class="" data-start="1038" data-end="1043" data-col-size="sm">模型</th>
<th class="" data-start="1043" data-end="1056" data-col-size="sm">VRAM需求（大約）</th>
<th class="" data-start="1056" data-end="1064" data-col-size="sm">推理速度</th>
</tr>
</thead>
<tbody data-start="1102" data-end="1165">
<tr data-start="1102" data-end="1120">
<td data-start="1102" data-end="1107" data-col-size="sm">2B</td>
<td data-col-size="sm" data-start="1107" data-end="1114">~4GB</td>
<td data-col-size="sm" data-start="1114" data-end="1120">很快</td>
</tr>
<tr data-start="1121" data-end="1142">
<td data-start="1121" data-end="1126" data-col-size="sm">7B</td>
<td data-col-size="sm" data-start="1126" data-end="1136">~8&ndash;16GB</td>
<td data-col-size="sm" data-start="1136" data-end="1142">中等</td>
</tr>
<tr data-start="1143" data-end="1165">
<td data-start="1143" data-end="1149" data-col-size="sm">27B</td>
<td data-col-size="sm" data-start="1149" data-end="1158">~40GB+</td>
<td data-col-size="sm" data-start="1158" data-end="1165">慢但強</td>
</tr>
</tbody>
</table>
</div>
</div>
<p data-start="1167" data-end="1175">👉 如果你用：</p>
<ul data-start="1176" data-end="1218">
<li data-section-id="4w522b" data-start="1176" data-end="1201">RTX 3060 / 4060 &rarr; 建議 7B</li>
<li data-section-id="qig6uz" data-start="1202" data-end="1218">無GPU &rarr; 2B + 量化</li>
</ul>
<hr data-start="1220" data-end="1223">
<h2 data-section-id="okhejh" data-start="1225" data-end="1242">6️⃣ 和其他模型的定位比較</h2>
<div class="TyagGW_tableContainer">
<div class="group TyagGW_tableWrapper flex flex-col-reverse w-fit" tabindex="-1">
<table class="w-fit min-w-(--thread-content-width)" data-start="1244" data-end="1484">
<thead data-start="1244" data-end="1255">
<tr data-start="1244" data-end="1255">
<th class="" data-start="1244" data-end="1249" data-col-size="sm">模型</th>
<th class="" data-start="1249" data-end="1255" data-col-size="sm">定位</th>
</tr>
</thead>
<tbody data-start="1272" data-end="1484">
<tr data-start="1272" data-end="1325">
<td data-start="1272" data-end="1312" data-col-size="sm"><span class="hover:entity-accent entity-underline inline cursor-pointer align-baseline"><span class="whitespace-normal">Gemma</span></span></td>
<td data-col-size="sm" data-start="1312" data-end="1325">開源友善、本地部署</td>
</tr>
<tr data-start="1326" data-end="1377">
<td data-start="1326" data-end="1366" data-col-size="sm"><span class="hover:entity-accent entity-underline inline cursor-pointer align-baseline"><span class="whitespace-normal">LLaMA</span></span></td>
<td data-col-size="sm" data-start="1366" data-end="1377">生態強、社群多</td>
</tr>
<tr data-start="1378" data-end="1430">
<td data-start="1378" data-end="1418" data-col-size="sm"><span class="hover:entity-accent entity-underline inline cursor-pointer align-baseline"><span class="whitespace-normal">Mistral</span></span></td>
<td data-col-size="sm" data-start="1418" data-end="1430">高效率、高性價比</td>
</tr>
<tr data-start="1431" data-end="1484">
<td data-start="1431" data-end="1471" data-col-size="sm"><span class="hover:entity-accent entity-underline inline cursor-pointer align-baseline"><span class="whitespace-normal">GPT</span></span></td>
<td data-col-size="sm" data-start="1471" data-end="1484">雲端最強（非開源）</td>
</tr>
</tbody>
</table>
</div>
</div>
<hr data-start="1486" data-end="1489">
<h1 data-section-id="xwaqk6" data-start="1491" data-end="1506">🔥 簡單選擇建議（重點）</h1>
<p data-start="1507" data-end="1514">直接給你結論：</p>
<ul data-start="1516" data-end="1638">
<li data-section-id="162qk3a" data-start="1516" data-end="1562">👉 <strong data-start="1521" data-end="1534">想本地跑 + 好用</strong><br>&rarr; Gemma 4 <strong data-start="1547" data-end="1562">7B Instruct</strong></li>
<li data-section-id="12csnc2" data-start="1564" data-end="1596">👉 <strong data-start="1569" data-end="1577">硬體很弱</strong><br>&rarr; Gemma 4 <strong data-start="1590" data-end="1596">2B</strong></li>
<li data-section-id="ergrah" data-start="1598" data-end="1638">👉 <strong data-start="1603" data-end="1618">追求效果（接近GPT）</strong><br>&rarr; Gemma 4 <strong data-start="1631" data-end="1638">27B</strong></li>
</ul>
<hr data-start="1640" data-end="1643">
<h1 data-section-id="10o8c9l" data-start="1645" data-end="1658">💡 補充（實務經驗）</h1>
<p data-start="1659" data-end="1672">Gemma 系列的特色是：</p>
<ul data-start="1674" data-end="1771">
<li data-section-id="18ph3p1" data-start="1674" data-end="1692">相對「乾淨」的訓練（安全性較高）</li>
<li data-section-id="rgsp7j" data-start="1693" data-end="1710">英文能力通常比中文好（需注意）</li>
<li data-section-id="z5z67j" data-start="1711" data-end="1771">很適合：
<ul data-start="1720" data-end="1771">
<li data-section-id="aq3b0" data-start="1720" data-end="1728">搭配 RAG</li>
<li data-section-id="z8wd4q" data-start="1731" data-end="1741">自建 AI 系統</li>
<li data-section-id="dnjkhp" data-start="1744" data-end="1771">本地部署（像你之前搞 open-webui 👍）</li>
</ul>
</li>
</ul>
<hr data-start="1773" data-end="1776">
<p data-start="1778" data-end="1864" data-is-last-node="" data-is-only-node="">如果你要，我可以幫你進一步分析👇<br data-start="1795" data-end="1798">👉「你的電腦規格適合跑哪一個 Gemma 4」<br data-start="1822" data-end="1825">👉 或「跟你現在用的模型（例如 open-webui 裡的）怎麼選最強組合」</p>