DeepSeek V3 與 V4 架構資訊圖
一份詳盡的並排技術資訊圖，比較 DeepSeek V3/R1 與 DeepSeek V4 的 Transformer 架構，適用於社群媒體貼文、簡報或模型分析視覺化。
提示詞
{"type":"並排 AI 架構比較資訊圖","style":"簡潔技術圖表，白色背景，細黑色輪廓，圓角矩形，虛線標註框，顏色編碼重點，簡報 Slides 美學，向量資訊圖","canvas":{"aspect_ratio":"2:1","resolution":"寬幅橫向"},"title_row":{"left_title":"DeepSeek V3/R1 (6,710 億參數)","right_title":"DeepSeek V4 (1.2 兆參數)","left_title_color":"亮橘紅色","right_title_color":"亮藍色"},"layout":{"columns":2,"sections":[{"title":"DeepSeek V3/R1 (6,710 億參數)","position":"左半部","count":9,"labels":["詞彙量 129k","FeedForward (SwiGLU) 模組","中間隱藏層維度 2,048","MoE 層","支援 128k token 的上下文長度","前 3 個區塊使用隱藏層大小為 18,432 的密集 FFN 而非 MoE","範例輸入文字","嵌入維度 7,168","128 個注意力頭"]},{"title":"DeepSeek V4 (1.2 兆參數)","position":"右半部","count":9,"labels":["詞彙量 160k","FeedForward (SwiGLU) 模組","中間隱藏層維度 3,072","MoE 層","支援 256k token 的上下文長度","前 3 個區塊使用隱藏層大小為 24,576 的密集 FFN 而非 MoE","範例輸入文字","嵌入維度 8,192","128 個注意力頭"]},{"title":"底部比較表","position":"底部全寬","count":10,"labels":["總參數","每個 token 的活躍參數","隱藏層大小","樣本設計","DeepSeek V3/R1","中間層 (FF)","注意力頭","上下文長度","嵌入維度","詞彙量"]}]},"left_panel":{"background":"淺灰色圓角矩形","main_stack":{"count":8,"blocks":["Token 化文字","Token 嵌入層","RMSNorm 1","多頭潛在注意力 (Multi-head Latent Attention)","RMSNorm 2","MoE","最終 RMSNorm","線性輸出層"]},"side_module":"RoPE 連接至左側注意力區塊","attention_block":{"label":"多頭潛在注意力 (Multi-head Latent Attention)","accent":"Latent 一詞使用橘紅色文字"},"feedforward_inset":{"title":"FeedForward (SwiGLU) 模組","count":4,"blocks":["線性層","SiLU 激活函數","線性層","線性層"],"diagram":"兩個分支相乘後進行投影"},"moe_inset":{"title":"MoE 層","count":5,"blocks":["頂部合併節點","前饋網路","前饋網路","路由器","專家數量標章 256"],"details":"小型黑色方塊，選取 1 位專家，箭頭指向專家，虛線分隔線"},"annotations":{"vocab":"詞彙量 129k","ff_dim":"中間隱藏層維度 2,048","context":"支援 128k token 的上下文長度","dense_first_blocks":"前 3 個區塊使用隱藏層大小為 18,432 的密集 FFN 而非 MoE","resource_savings":"資源節省：模型大小為 671B，但每個 token 僅啟動 1 個（共享）+ 8 個專家；每次推理步驟僅啟動 37B 參數"},"bottom_stats":{"count":10,"items":["總參數：671B","每個 token 活躍參數：37B (1 + 8 位專家)","隱藏層大小：7,128","樣本設計：28,432","中間層 (FF)：2,048","注意力頭：128","上下文長度：128k","嵌入維度：前 3 個區塊","上下文長度：22G7","詞彙量：129k"]}},"right_panel":{"background":"淺藍色圓角矩形","main_stack":{"count":8,"blocks":["Token 化文字","Token 嵌入層","RMSNorm 1","多頭潛在注意力 (Multi-head Latent Attention)","RMSNorm 2","MoE","最終 RMSNorm","線性輸出層"]},"side_module":"RoPE 連接至左側注意力區塊","attention_block":{"label":"多頭潛在注意力 (Multi-head Latent Attention)","accent":"Latent 一詞使用藍色文字"},"feedforward_inset":{"title":"FeedForward (SwiGLU) 模組","count":4,"blocks":["線性層","SiLU 激活函數","線性層","線性層"],"diagram":"與左側面板結構相同"},"moe_inset":{"title":"MoE 層","count":5,"blocks":["頂部合併節點","前饋網路","前饋網路","路由器","專家數量標章 384"],"details":"小型黑色方塊，選取 1 位專家，箭頭指向專家，虛線分隔線，藍色邊框強調"},"annotations":{"vocab":"詞彙量 160k","ff_dim":"中間隱藏層維度 3,072","context":"支援 256k token 的上下文長度","dense_first_blocks":"前 3 個區塊使用隱藏層大小為 24,576 的密集 FFN 而非 MoE","resource_savings":"資源節省：模型大小為 1.2T，但每個 token 僅啟動 1 個（共享）+ 8 個專家；每次推理步驟僅啟動 52B 參數"},"bottom_stats":{"count":10,"items":["總參數：1.2T","每個 token 活躍參數：52B (1 + 8 位專家)","隱藏層大小：7,2B","樣本設計：28,432","中間層 (FF)：3,072","注意力頭：128","上下文長度：256k","嵌入維度：前 3 個區塊","上下文長度：22G7","詞彙量：160k"]}},"global_notes":"建立一份極其詳細的 Transformer 架構比較圖，採用鏡像佈局。每一半包含一個大型模型堆疊圖以及 2 個嵌入圖：1 個前饋模組和 1 個 MoE 層。在區塊之間使用箭頭、微小的技術標籤，並從標籤連接至相關組件。保持排版緊湊且具備 Slides 風格，V3/R1 的重點統一使用橘紅色，V4 的重點統一使用藍色。包含底部一列橫跨寬度的緊湊表格指標。保留略帶手工感、資訊密度高且標註擁擠的資訊圖風格。"}
如何使用這條提示詞
- 1
複製上方完整的提示詞。
- 2
開啟支援 GPT Image 2 的平台（例如 YouMind），把提示詞貼上。
- 3
依你的想法替換主體、風格或細節，然後生成。
這是 YouMind 提示詞庫裡的一條免費 AI 提示詞。這裡還有成千上萬條 圖像 提示詞，都能免費複製與改用。
探索更多 圖像 提示詞
更多提示詞功能
