{
  "version": "https://jsonfeed.org/version/1", 
  "title": "\u7845\u57fa\u751f\u6d3b", 
  "description": "- RTX 4090|5090 >24g / MBP|M5u|DgxSpark >128g\r\n- Claude/OpenAI/Deepseek\r\n- Grokbot/Muse/Dots/Duet\r\n- Suno/Seedance/Minimax/ElevenLabs/REA", 
  "home_page_url": "https://www.v2ex.com/go/siliconlife", 
  "feed_url": "https://www.v2ex.com/feed/siliconlife.json", 
  "icon": "https://cdn.v2ex.com/navatar/d821/1837/1381_large.png?m=1794242127", 
  "favicon": "https://cdn.v2ex.com/navatar/d821/1837/1381_normal.png?m=1794242127", 
  "items": [
    {
      "author": {
        "url": "https://www.v2ex.com/member/davidqw", 
        "name": "davidqw", 
        "avatar": "https://cdn.v2ex.com/avatar/d847/a023/15793_large.png?m=1770803455"
      }, 
      "url": "https://www.v2ex.com/t/1255702", 
      "date_modified": "2026-11-09T16:50:08+00:00", 
      "content_html": "<p>\u5355\u53f0 Mac Studio M5U 512G \u53ef\u4ee5\u5b8c\u6574\u8fd0\u884c dsf v4.1 \uff0c\u800c\u4e14\u80fd\u8dd1\u6743\u91cd\u4e0e\u5b98\u65b9\u5b8c\u5168\u4e00\u81f4\u7684\u7248\u672c\u3002\u6700\u4f73\u7ec4\u5408\u662f\uff1aoMLX 0.7.3 + <a href=\"https://huggingface.co/mlx-community/DeepSeek-V4.1-Flash-1017-oQ4e-ngram-mtp\" rel=\"nofollow\">oQ4e-ngram-mtp</a>\uff08\u5e38\u9a7b\u7ea6 288 GiB \uff0cEngram \u8868\u653e SSD \u6309\u9700\u8bfb\u53d6\uff09\uff0c\u6253\u5f00 DSpark \u6295\u673a\u89e3\u7801\u3001CED prefill \u548c\u524d\u7f00\u7f13\u5b58\u3002</p>\n<table>\n<thead>\n<tr>\n<th>\u9879\u76ee</th>\n<th>\u9009\u62e9</th>\n<th>\u7406\u7531</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>\u63a8\u7406\u5f15\u64ce</td>\n<td>oMLX 0.7.3 \u6216\u66f4\u65b0\uff08\u56fa\u5b9a\u7248\u672c\uff09</td>\n<td>\u5b8c\u6574\u652f\u6301 V4.1 \u7684 Engram \u3001CSA2 \u3001DSpark \uff1b prefill \u5df2\u7528\u4e0a M5 \u7684 GPU \u5f20\u91cf\u5355\u5143</td>\n</tr>\n<tr>\n<td>\u6a21\u578b</td>\n<td>mlx-community/DeepSeek-V4.1-Flash-1017-oQ4e-ngram-mtp</td>\n<td>\u4e13\u5bb6\u3001\u4e3b\u5e72\u548c Engram \u90fd\u4fdd\u6301\u5b98\u65b9\u539f\u751f\u7cbe\u5ea6\uff0c\u6ca1\u6709\u989d\u5916\u91cf\u5316\uff1b\u5305\u542b DSpark \u6743\u91cd</td>\n</tr>\n<tr>\n<td>Engram \u8868</td>\n<td>\u653e\u5185\u7f6e SSD \uff0c\u6309\u9700\u8bfb\u53d6</td>\n<td>\u6bcf\u4e2a token \u6bcf\u5f20\u8868\u53ea\u8bfb\u4e00\u884c\uff0c\u51e0\u4e4e\u4e0d\u5f71\u54cd decode \uff1b\u7701\u4e0b\u7ea6 189 GiB \u5185\u5b58</td>\n</tr>\n<tr>\n<td>\u6295\u673a\u89e3\u7801</td>\n<td>DSpark \uff0c\u6bcf\u8f6e\u8349\u62df 5 \u4e2a token</td>\n<td>\u5355\u673a\u6700\u5927\u7684\u63d0\u901f\u624b\u6bb5</td>\n</tr>\n<tr>\n<td>Prefill</td>\n<td>CED prefill + INT8 Activation Prefill</td>\n<td>1M \u4e0a\u4e0b\u6587\u65f6 prefill \u65f6\u95f4\u662f\u4e3b\u8981\u74f6\u9888</td>\n</tr>\n<tr>\n<td>\u7f13\u5b58</td>\n<td>\u524d\u7f00\u7f13\u5b58 + SSD \u5feb\u7167</td>\n<td>agent \u591a\u8f6e\u5bf9\u8bdd\u4e0d\u7528\u6bcf\u6b21\u91cd\u7b97\u6574\u6bb5\u4e0a\u4e0b\u6587</td>\n</tr>\n</tbody></table><p>\u4e0a\u4e0b\u6587\u6309 1M token \u89c4\u5212\u3002\u5185\u5b58\u4f59\u91cf\u7ea6 180 GiB \uff0c\u5b8c\u5168\u591f\u7528\uff1b\u51b7\u542f\u52a8 prefill \u7ea6\u9700 6\u20138 \u5206\u949f\uff0c\u914d\u7f6e\u548c\u7528\u6cd5\u89c1\u7b2c 6 \u8282\u3002</p>\n<p>\u9884\u671f\u6027\u80fd\u5982\u4e0b\u3002prefill \u548c\u5e76\u53d1\u6570\u636e\u6765\u81ea oMLX 0.7.3 \u5728 M5 Ultra \u4e0a\u7684\u5b9e\u6d4b\uff08\u53d1\u5e03\u8bf4\u660e\u672a\u6ce8\u660e\u91cf\u5316\u7248\u672c\uff09\uff1b<strong>\u5355\u8def decode \u4ecd\u662f\u63a8\u7b97\u503c</strong>\uff0c\u62ff\u5230\u673a\u5668\u540e\u7528\u7b2c 8 \u8282\u7684\u6d41\u7a0b\u66ff\u6362\u6210\u5b9e\u6d4b\u503c\u3002</p>\n<table>\n<thead>\n<tr>\n<th>\u6307\u6807</th>\n<th>\u6570\u503c</th>\n<th>\u4f9d\u636e</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>decode \uff0c\u4e0d\u5f00 DSpark</td>\n<td>\u7ea6 55\u201360 tok/s \uff08\u63a8\u7b97\uff09</td>\n<td>M5 Ultra \u5b9e\u6d4b\uff1a3.0bpw \u7248\u672c 67\u201372 tok/s \uff08 mlx-serve \uff09\uff1b 4-bit \u6bcf token \u591a\u8bfb\u7ea6 20% \u6743\u91cd</td>\n</tr>\n<tr>\n<td>decode \uff0c\u5f00 DSpark</td>\n<td>\u7ea6 80\u2013110 tok/s \uff08\u63a8\u7b97\uff0c\u4ee3\u7801\u504f\u9ad8\uff09</td>\n<td>\u540c\u4e00\u5b9e\u6d4b\uff1a3.0bpw \u7248\u672c\u5f00 DSpark 96\u2013126 tok/s</td>\n</tr>\n<tr>\n<td>prefill \uff0c128K</td>\n<td>2,390 tok/s \uff08\u5b9e\u6d4b\uff09</td>\n<td>oMLX 0.7.3 \uff0c\u5f00 Engram \u9884\u53d6</td>\n</tr>\n<tr>\n<td>prefill \uff0c64K / 195K</td>\n<td>3,507 / 3,200 tok/s \uff08\u5b9e\u6d4b\uff09</td>\n<td>oMLX 0.7.3 \uff0c\u5f00 INT8 Activation Prefill</td>\n</tr>\n<tr>\n<td>\u51b7\u542f\u52a8 TTFT \uff0c8K / 128K / 1M</td>\n<td>\u7ea6 3 \u79d2 / 45\u201355 \u79d2 / 6\u20138 \u5206\u949f</td>\n<td>\u7531\u4e0a\u4e24\u884c\u63a8\u7b97</td>\n</tr>\n<tr>\n<td>12 \u4e2a\u5e76\u53d1\u8bf7\u6c42\u5408\u8ba1 decode</td>\n<td>125.7 tok/s \uff08\u5b9e\u6d4b\uff09</td>\n<td>oMLX 0.7.3</td>\n</tr>\n<tr>\n<td>\u5cf0\u503c\u529f\u8017</td>\n<td>\u7ea6 300 W</td>\n<td>Apple \u8bba\u6587\u5b9e\u6d4b</td>\n</tr>\n</tbody></table><h2>\u6a21\u578b\u7279\u70b9\u4e0e\u5185\u5b58\u9884\u7b97</h2>\n<p>oQ4e-ngram-mtp \u5171 476.8 GiB \uff0c\u5176\u4e2d Engram \u8868 188.8 GiB \u7559\u5728 SSD \u4e0a\u6309\u9700\u8bfb\u53d6\uff0c\u5e38\u9a7b\u5185\u5b58\u7ea6 288 GiB \u3002\u628a GPU \u53ef\u7528\u5185\u5b58\u4e0a\u9650\u8c03\u5230\u7ea6 469 GiB \u540e\uff0c\u8fd8\u5269\u7ea6 180 GiB \u7ed9 KV cache \u3001prefill \u7f13\u51b2\u3001\u5e76\u53d1\u4f1a\u8bdd\u548c MLX \u7f13\u5b58\u3002</p>\n<table>\n<thead>\n<tr>\n<th>\u7ec4\u4ef6</th>\n<th>\u5b58\u50a8\u5927\u5c0f</th>\n<th>\u4f4d\u7f6e</th>\n<th>\u8bf4\u660e</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>\u8bed\u8a00\u4e3b\u5e72</td>\n<td>279.5 GiB</td>\n<td>\u5185\u5b58</td>\n<td>\u8def\u7531\u4e13\u5bb6 MXFP4 \uff0c\u6ce8\u610f\u529b\u548c\u7a20\u5bc6\u5c42 MXFP8 \uff0c\u90fd\u662f\u5b98\u65b9\u539f\u751f\u7cbe\u5ea6</td>\n</tr>\n<tr>\n<td>DSpark \u5934</td>\n<td>7.5 GiB</td>\n<td>\u5185\u5b58</td>\n<td>\u6295\u673a\u89e3\u7801\u8981\u7528</td>\n</tr>\n<tr>\n<td>\u89c6\u89c9\u7f16\u7801\u5668</td>\n<td>0.9 GiB</td>\n<td>\u5185\u5b58</td>\n<td>\u56fe\u7247\u8f93\u5165\u8981\u7528</td>\n</tr>\n<tr>\n<td>Engram \u8868</td>\n<td>188.8 GiB</td>\n<td>\u5185\u7f6e SSD</td>\n<td>\u4fdd\u6301\u5b98\u65b9 FP8 \uff08 E4M3 + \u6bcf 32 \u901a\u9053\u4e00\u4e2a E8M0 scale \uff09\uff0c\u65e0\u635f</td>\n</tr>\n<tr>\n<td><strong>\u5408\u8ba1</strong></td>\n<td><strong>476.8 GiB</strong></td>\n<td>\u5e38\u9a7b\u7ea6 288 GiB</td>\n<td>\u4e0b\u8f7d\u5927\u5c0f 512 GB</td>\n</tr>\n</tbody></table><p>\u548c\u90e8\u7f72\u65b9\u5f0f\u76f4\u63a5\u76f8\u5173\u7684\u51e0\u4e2a\u67b6\u6784\u7279\u70b9\uff1a</p>\n<ul>\n<li><strong>MoE</strong>\uff1a552B \u4e3b\u5e72\u53c2\u6570\uff0c\u6bcf\u5c42 384 \u4e2a\u8def\u7531\u4e13\u5bb6\u6fc0\u6d3b 6 \u4e2a\uff0c\u53e6\u6709 1 \u4e2a\u5171\u4eab\u4e13\u5bb6\u3002decode \u65f6\u6bcf\u4e2a token \u6fc0\u6d3b 16B \uff0c\u901f\u5ea6\u4e3b\u8981\u53d7\u5185\u5b58\u5e26\u5bbd\u9650\u5236\u3002</li>\n<li><strong>Causal Encoder-Decoder \uff08 CED \uff09</strong>\uff1a40 \u5c42\u5206\u6210 20 \u5c42 encoder \u548c 20 \u5c42 decoder \u3002decoder \u7684\u5168\u5c40 KV \u7531 encoder \u6700\u540e\u4e00\u5c42\u6295\u5f71\u5f97\u5230\uff0c\u6240\u4ee5 prefill \u53ea\u9700\u6fc0\u6d3b 8B \u3002\u8fd9\u5c31\u662f CED prefill \u4f18\u5316\u7684\u524d\u63d0\u3002</li>\n<li><strong>CSA2 \u7a00\u758f\u6ce8\u610f\u529b</strong>\uff1a\u53ea\u6709 4 \u5c42\u8ba1\u7b97\u81ea\u5df1\u7684 KV \uff0c\u5176\u4f59\u5c42\u5171\u7528\u3002KV cache \u53ea\u6709 890 \u5b57\u8282/token \uff0c1M \u4e0a\u4e0b\u6587\u7684\u5168\u5c40 KV \u4e5f\u53ea\u6709\u7ea6 0.83 GiB \u3002<strong>\u4e0a\u4e0b\u6587\u957f\u5ea6\u51e0\u4e4e\u4e0d\u53d7\u5185\u5b58\u9650\u5236\uff0c\u53d7\u7684\u662f prefill \u65f6\u95f4\u9650\u5236\u3002</strong></li>\n<li><strong>Engram n-gram \u8bb0\u5fc6</strong>\uff1a\u7b2c 1 \u5c42\u548c\u7b2c 14 \u5c42\u5404\u6709\u4e00\u5f20\u7ea6 3.84 \u4ebf\u884c \u00d7 256 \u7ef4\u7684\u54c8\u5e0c\u8868\u3002\u6bcf\u4e2a token \u53ea\u67e5\u51e0\u884c\uff0c<strong>\u53ef\u4ee5\u653e\u5728 SSD \u4e0a\u6309\u9700\u8bfb\u53d6</strong>\u3002</li>\n<li><strong>Hyper-Connections</strong>\uff1a\u6b8b\u5dee\u6d41\u6709 4 \u8def\u526f\u672c\uff0c\u6bcf\u4e2a\u5b50\u5c42\u505a 20 \u6b21 Sinkhorn \u8fed\u4ee3\u3002\u5c0f\u7b97\u5b50\u5f88\u591a\uff0c\u6240\u4ee5\u5f15\u64ce\u6709\u6ca1\u6709\u505a\u7b97\u5b50\u878d\u5408\u5bf9\u901f\u5ea6\u5f71\u54cd\u5f88\u5927\u3002</li>\n<li><strong>DSpark</strong>\uff1a\u5171 3 \u7ea7\uff0c\u6bcf\u7ea7 128 \u4e2a\u4e13\u5bb6\u6fc0\u6d3b 3 \u4e2a\uff0c\u8bfb\u4e3b\u6a21\u578b\u7b2c 37\u201339 \u5c42\u7684\u8f93\u5165\uff0c\u6bcf\u8f6e\u8349\u62df 5 \u4e2a token \u3002\u5b83\u548c\u4e3b\u6a21\u578b\u7ed1\u5728\u4e00\u8d77\uff0c\u5fc5\u987b\u5728\u540c\u4e00\u53f0\u673a\u5668\u4e0a\u8dd1\u3002</li>\n</ul>\n", 
      "date_published": "2026-11-09T16:49:23+00:00", 
      "title": "\u672c\u5730\u90e8\u7f72\u539f\u751f\u7cbe\u5ea6 DeepSeek-V4.1-Flash-1017 \u65b9\u6848", 
      "id": "https://www.v2ex.com/t/1255702"
    }
  ]
}