<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[模型卡里的上下文长度，和实际可用长度不是一回事]]></title><description><![CDATA[<p dir="auto">模型卡写 128K 或更长上下文，是不是代表我可以放心塞 120K token 资料，答案还稳定？</p>
]]></description><link>https://localaihub.com/topic/99/模型卡里的上下文长度-和实际可用长度不是一回事</link><generator>RSS for Node</generator><lastBuildDate>Wed, 03 Jun 2026 17:50:57 GMT</lastBuildDate><atom:link href="https://localaihub.com/topic/99.rss" rel="self" type="application/rss+xml"/><pubDate>Wed, 06 May 2026 08:35:00 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Thu, 07 May 2026 09:26:00 GMT]]></title><description><![CDATA[<p dir="auto">长上下文是能力，不是清洁工。</p>
]]></description><link>https://localaihub.com/post/768</link><guid isPermaLink="true">https://localaihub.com/post/768</guid><dc:creator><![CDATA[小高]]></dc:creator><pubDate>Thu, 07 May 2026 09:26:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Thu, 07 May 2026 07:25:00 GMT]]></title><description><![CDATA[<p dir="auto">对。窗口越大，越需要整理输入。</p>
]]></description><link>https://localaihub.com/post/767</link><guid isPermaLink="true">https://localaihub.com/post/767</guid><dc:creator><![CDATA[nora]]></dc:creator><pubDate>Thu, 07 May 2026 07:25:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Thu, 07 May 2026 04:59:00 GMT]]></title><description><![CDATA[<p dir="auto">所以窗口大不代表上下文管理可以摆烂。</p>
]]></description><link>https://localaihub.com/post/766</link><guid isPermaLink="true">https://localaihub.com/post/766</guid><dc:creator><![CDATA[小傅]]></dc:creator><pubDate>Thu, 07 May 2026 04:59:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Thu, 07 May 2026 03:48:00 GMT]]></title><description><![CDATA[<p dir="auto">我们之前把政策合集按月份拼一起，模型总引用旧政策。后来按生效日期过滤，效果才好。</p>
]]></description><link>https://localaihub.com/post/765</link><guid isPermaLink="true">https://localaihub.com/post/765</guid><dc:creator><![CDATA[小满]]></dc:creator><pubDate>Thu, 07 May 2026 03:48:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Thu, 07 May 2026 01:00:00 GMT]]></title><description><![CDATA[<p dir="auto">还要测随着长度增加的成本和延迟。用户不只要答对，还要等得起。</p>
]]></description><link>https://localaihub.com/post/764</link><guid isPermaLink="true">https://localaihub.com/post/764</guid><dc:creator><![CDATA[林小北]]></dc:creator><pubDate>Thu, 07 May 2026 01:00:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Wed, 06 May 2026 23:39:00 GMT]]></title><description><![CDATA[<p dir="auto">做位置敏感测试：关键答案放开头、中间、末尾；单文档、多文档；有干扰段落；看引用是否正确。</p>
]]></description><link>https://localaihub.com/post/763</link><guid isPermaLink="true">https://localaihub.com/post/763</guid><dc:creator><![CDATA[zeroOne]]></dc:creator><pubDate>Wed, 06 May 2026 23:39:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Wed, 06 May 2026 22:41:00 GMT]]></title><description><![CDATA[<p dir="auto">怎么测实际可用长度？</p>
]]></description><link>https://localaihub.com/post/762</link><guid isPermaLink="true">https://localaihub.com/post/762</guid><dc:creator><![CDATA[小周]]></dc:creator><pubDate>Wed, 06 May 2026 22:41:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Wed, 06 May 2026 20:29:00 GMT]]></title><description><![CDATA[<p dir="auto">如果真实体验没测过，别这么写。可以写“支持长文档处理”，但验收要看准确率和延迟。</p>
]]></description><link>https://localaihub.com/post/761</link><guid isPermaLink="true">https://localaihub.com/post/761</guid><dc:creator><![CDATA[葡萄冰]]></dc:creator><pubDate>Wed, 06 May 2026 20:29:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Wed, 06 May 2026 18:45:00 GMT]]></title><description><![CDATA[<p dir="auto">那产品宣传能写“支持 128K 文档问答”吗？</p>
]]></description><link>https://localaihub.com/post/760</link><guid isPermaLink="true">https://localaihub.com/post/760</guid><dc:creator><![CDATA[普通网友A]]></dc:creator><pubDate>Wed, 06 May 2026 18:45:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Wed, 06 May 2026 17:02:00 GMT]]></title><description><![CDATA[<p dir="auto">API 文档里的限制也要看输入输出合计、单次请求限制、模型版本。别用旧博客数字。</p>
]]></description><link>https://localaihub.com/post/759</link><guid isPermaLink="true">https://localaihub.com/post/759</guid><dc:creator><![CDATA[leaf_1997]]></dc:creator><pubDate>Wed, 06 May 2026 17:02:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Wed, 06 May 2026 16:33:00 GMT]]></title><description><![CDATA[<p dir="auto">本地部署还受显存、KV cache、推理框架限制。模型理论窗口和你实际能跑的窗口可能不同。</p>
]]></description><link>https://localaihub.com/post/758</link><guid isPermaLink="true">https://localaihub.com/post/758</guid><dc:creator><![CDATA[rootless]]></dc:creator><pubDate>Wed, 06 May 2026 16:33:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Wed, 06 May 2026 13:26:00 GMT]]></title><description><![CDATA[<p dir="auto">还有输出 token 要留空间。你塞满输入，模型没地方回答。</p>
]]></description><link>https://localaihub.com/post/757</link><guid isPermaLink="true">https://localaihub.com/post/757</guid><dc:creator><![CDATA[陈一]]></dc:creator><pubDate>Wed, 06 May 2026 13:26:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Wed, 06 May 2026 11:41:00 GMT]]></title><description><![CDATA[<p dir="auto">长上下文模型在不同位置找信息的能力会波动，尤其多文档混杂时。LongBench 这类评测就是提醒大家别只看窗口数字。</p>
]]></description><link>https://localaihub.com/post/756</link><guid isPermaLink="true">https://localaihub.com/post/756</guid><dc:creator><![CDATA[小高]]></dc:creator><pubDate>Wed, 06 May 2026 11:41:00 GMT</pubDate></item><item><title><![CDATA[Reply to 模型卡里的上下文长度，和实际可用长度不是一回事 on Wed, 06 May 2026 10:27:00 GMT]]></title><description><![CDATA[<p dir="auto">不是。上下文长度是窗口容量，不是全窗口等质量注意力保证。</p>
]]></description><link>https://localaihub.com/post/755</link><guid isPermaLink="true">https://localaihub.com/post/755</guid><dc:creator><![CDATA[nora]]></dc:creator><pubDate>Wed, 06 May 2026 10:27:00 GMT</pubDate></item></channel></rss>