官方模型技术参数
直接来自每个模型的 config.json，不是二手数据。
这些参数决定能不能跑、要多少显存、上下文多长。
几个值得注意的。
V4 系列的
max_position_embeddings 是 1,048,576 —— 这就是「百万上下文」的来源，
不是宣传口径而是配置里的实际值。V4-Flash 只有 43 层，比 V4-Pro 的 61 层小，所以便宜。| 模型 | 层数 | 隐藏维度 | 上下文 | 词表 | 结构 |
|---|---|---|---|---|---|
| DeepSeek-V4-Flash | 43 | 4,096 | 1,048,576（1M） | 129,280 | MoE |
| DeepSeek-V4-Flash-0731 | 43 | 4,096 | 1,048,576（1M） | 129,280 | MoE |
| DeepSeek-V4-Flash-Base | 43 | 4,096 | 1,048,576（1M） | 129,280 | MoE |
| DeepSeek-V4-Flash-DSpark | 43 | 4,096 | 1,048,576（1M） | 129,280 | MoE |
| DeepSeek-V4-Flash-Vision-Exp | 43 | 4,096 | 1,048,576（1M） | 129,280 | MoE |
| DeepSeek-V4-Pro | 61 | 7,168 | 1,048,576（1M） | 129,280 | MoE |
| DeepSeek-V4-Pro-0813 | 61 | 7,168 | 1,048,576（1M） | 129,280 | MoE |
| DeepSeek-V4-Pro-Base | 61 | 7,168 | 1,048,576（1M） | 129,280 | MoE |
| DeepSeek-V4-Pro-DSpark | 61 | 7,168 | 1,048,576（1M） | 129,280 | MoE |
| dspark_gemma4_12b_block7 | 5 | 3,840 | 262,144 | 262,144 | Dense |
| DeepSeek-Coder-V2-Base | 60 | 5,120 | 163,840 | 102,400 | MoE |
| DeepSeek-Coder-V2-Instruct | 60 | 5,120 | 163,840 | 102,400 | MoE |
| DeepSeek-Coder-V2-Instruct-0724 | 60 | 5,120 | 163,840 | 102,400 | MoE |
| DeepSeek-Coder-V2-Lite-Base | 27 | 2,048 | 163,840 | 102,400 | MoE |
| DeepSeek-Coder-V2-Lite-Instruct | 27 | 2,048 | 163,840 | 102,400 | MoE |
| DeepSeek-Math-V2 | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-Prover-V2-671B | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-R1 | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-R1-0528 | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-R1-Zero | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-V2 | 60 | 5,120 | 163,840 | 102,400 | MoE |
| DeepSeek-V2-Chat | 60 | 5,120 | 163,840 | 102,400 | MoE |
| DeepSeek-V2-Chat-0628 | 60 | 5,120 | 163,840 | 102,400 | MoE |
| DeepSeek-V2-Lite | 27 | 2,048 | 163,840 | 102,400 | MoE |
| DeepSeek-V2-Lite-Chat | 27 | 2,048 | 163,840 | 102,400 | MoE |
| DeepSeek-V2.5 | 60 | 5,120 | 163,840 | 102,400 | MoE |
| DeepSeek-V2.5-1210 | 60 | 5,120 | 163,840 | 102,400 | MoE |
| DeepSeek-V3 | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-V3-0324 | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-V3-Base | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-V3.1 | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-V3.1-Base | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-V3.1-Terminus | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-V3.2 | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-V3.2-Exp | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-V3.2-Exp-Base | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-V3.2-Speciale | 61 | 7,168 | 163,840 | 129,280 | MoE |
| DeepSeek-R1-0528-Qwen3-8B | 36 | 4,096 | 131,072 | 151,936 | Dense |
| DeepSeek-R1-Distill-Llama-70B | 80 | 8,192 | 131,072 | 128,256 | Dense |
| DeepSeek-R1-Distill-Llama-8B | 32 | 4,096 | 131,072 | 128,256 | Dense |
| DeepSeek-R1-Distill-Qwen-1.5B | 28 | 1,536 | 131,072 | 151,936 | Dense |
| DeepSeek-R1-Distill-Qwen-14B | 48 | 5,120 | 131,072 | 152,064 | Dense |
| DeepSeek-R1-Distill-Qwen-32B | 64 | 5,120 | 131,072 | 152,064 | Dense |
| DeepSeek-R1-Distill-Qwen-7B | 28 | 3,584 | 131,072 | 152,064 | Dense |
| DeepSeek-Prover-V2-7B | 30 | 4,096 | 65,536 | 102,400 | Dense |
| dflash_qwen3_4b_block7 | 5 | 2,560 | 40,960 | 151,936 | Dense |
| dflash_qwen3_8b_block7 | 5 | 4,096 | 40,960 | 151,936 | Dense |
| dspark_qwen3_14b_block7 | 5 | 5,120 | 40,960 | 151,936 | Dense |
| dspark_qwen3_4b_block7 | 5 | 2,560 | 40,960 | 151,936 | Dense |
| dspark_qwen3_8b_block7 | 5 | 4,096 | 40,960 | 151,936 | Dense |
| deepseek-coder-1.3b-base | 24 | 2,048 | 16,384 | 32,256 | Dense |
| deepseek-coder-1.3b-instruct | 24 | 2,048 | 16,384 | 32,256 | Dense |
| deepseek-coder-33b-base | 62 | 7,168 | 16,384 | 32,256 | Dense |
| deepseek-coder-33b-instruct | 62 | 7,168 | 16,384 | 32,256 | Dense |
| deepseek-coder-5.7bmqa-base | 32 | 4,096 | 16,384 | 32,256 | Dense |
| deepseek-coder-6.7b-base | 32 | 4,096 | 16,384 | 32,256 | Dense |
| deepseek-coder-6.7b-instruct | 32 | 4,096 | 16,384 | 32,256 | Dense |
| DeepSeek-OCR | 12 | 1,280 | 8,192 | 129,280 | MoE |
| DeepSeek-OCR-2 | 12 | 1,280 | 8,192 | 129,280 | MoE |
| DeepSeek-Prover-V1 | 30 | 4,096 | 4,096 | 102,400 | Dense |
| DeepSeek-Prover-V1.5-Base | 30 | 4,096 | 4,096 | 102,400 | Dense |
| DeepSeek-Prover-V1.5-RL | 30 | 4,096 | 4,096 | 102,400 | Dense |
| DeepSeek-Prover-V1.5-SFT | 30 | 4,096 | 4,096 | 102,400 | Dense |
| ESFT-gate-code-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-gate-intent-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-gate-law-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-gate-math-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-gate-summary-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-gate-translation-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-token-code-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-token-intent-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-token-law-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-token-math-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-token-summary-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-token-translation-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| ESFT-vanilla-lite | 27 | 2,048 | 4,096 | 102,400 | MoE |
| deepseek-coder-7b-base-v1.5 | 30 | 4,096 | 4,096 | 102,400 | Dense |
| deepseek-coder-7b-instruct-v1.5 | 30 | 4,096 | 4,096 | 102,400 | Dense |
| deepseek-llm-67b-base | 95 | 8,192 | 4,096 | 102,400 | Dense |
| deepseek-llm-67b-chat | 95 | 8,192 | 4,096 | 102,400 | Dense |
| deepseek-llm-7b-base | 30 | 4,096 | 4,096 | 102,400 | Dense |
| deepseek-llm-7b-chat | 30 | 4,096 | 4,096 | 102,400 | Dense |
| deepseek-math-7b-base | 30 | 4,096 | 4,096 | 102,400 | Dense |
| deepseek-math-7b-instruct | 30 | 4,096 | 4,096 | 102,400 | Dense |
| deepseek-math-7b-rl | 30 | 4,096 | 4,096 | 102,400 | Dense |
| deepseek-moe-16b-base | 28 | 2,048 | 4,096 | 102,400 | MoE |
| deepseek-moe-16b-chat | 28 | 2,048 | 4,096 | 102,400 | MoE |
怎么用这张表
| 你要判断 | 看哪列 |
|---|---|
| 显存够不够 | 层数 × 隐藏维度（越大越吃显存） |
| 能不能塞下长文档 | 上下文 |
| 是不是 MoE（稀疏激活） | 结构列 |
| 词表大小影响嵌入层 | 词表 |
说明
数据来自 Hugging Face 模型卡里的
config.json，每周刷新。
参数量（总参数/激活参数）官方公布的口径与配置不同，这里只给配置里的确定值。