<![CDATA[Qwen3.8-27B-MLX-8bit on M5 max]]>Qwen3.8发布，先试了试在M5 Max上跑的速度，用的是lmstudio-community/Qwen3.8-27B-MLX-8bit，结果如下，供大家参考。
截屏2026-08-15 上午12.29.38.png

]]>https://lcz.me/topic/1123RSS for NodeMon, 07 Sep 2026 18:59:30 GMTFri, 14 Aug 2026 16:32:49 GMT60<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Mon, 31 Aug 2026 22:54:55 GMT]]>沒研究過Apples , 放在供大家 參考一下價格

Apple "MacBook Pro" M5 Max 基本款 約14萬台幣 , 18C CPU/32C GPU , 36 GB Unified Memory,

f2ce8c96-6807-40fd-9b06-402b90b65926-image.jpeg

Apple "MacBook Pro" M5 Max 頂規 約24萬台幣 , 18C CPU/40C GPU , 128 GB Unified Memory,

9fcd9c84-32e4-4f78-90a3-cf547c43162c-image.jpeg

]]>https://lcz.me/post/15213https://lcz.me/post/15213Mon, 31 Aug 2026 22:54:55 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Mon, 31 Aug 2026 22:44:02 GMT]]>

@zorg said:

开了MTP Q6可以跑到40-50tps，还是多模态无审查模型。

如果是 MTP Q4 不就可以上50-60 tps, 但要看輸出品質的差異是否可用？
Agent 可能差異不大 但coding或許有差異

]]>https://lcz.me/post/15212https://lcz.me/post/15212Mon, 31 Aug 2026 22:44:02 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Mon, 31 Aug 2026 08:10:02 GMT]]>@zorg 还行，prefill如何，如果不慢这速度够用，加上苹果的安静，是有市场的。

]]>https://lcz.me/post/15101https://lcz.me/post/15101Mon, 31 Aug 2026 08:10:02 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Mon, 31 Aug 2026 05:32:13 GMT]]>@zorg 这个速度感觉真能用了。就是价格还是有点美丽😢

]]>https://lcz.me/post/15080https://lcz.me/post/15080Mon, 31 Aug 2026 05:32:13 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Mon, 31 Aug 2026 08:09:17 GMT]]>@terry
截屏2026-08-31 下午1.14.29.png

]]>https://lcz.me/post/15078https://lcz.me/post/15078Mon, 31 Aug 2026 08:09:17 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 30 Aug 2026 21:59:26 GMT]]>@zorg 那还挺好，长上下文如何，我看最高才32K，意义不大。max的算力还是不错啊，其实640G的带宽和R9700 Pro一样了。

]]>https://lcz.me/post/15017https://lcz.me/post/15017Sun, 30 Aug 2026 21:59:26 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 30 Aug 2026 20:33:24 GMT]]>开了MTP Q6可以跑到40-50tps，还是多模态无审查模型。纯文本Deepseek V4 flash也是可用的速度。

]]>https://lcz.me/post/15014https://lcz.me/post/15014Sun, 30 Aug 2026 20:33:24 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 30 Aug 2026 20:30:35 GMT]]>DEEPSEEKV4Flash0731.png

]]>https://lcz.me/post/15012https://lcz.me/post/15012Sun, 30 Aug 2026 20:30:35 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 30 Aug 2026 20:30:11 GMT]]>截屏2026-08-31 上午4.21.21.png

]]>https://lcz.me/post/15011https://lcz.me/post/15011Sun, 30 Aug 2026 20:30:11 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 30 Aug 2026 20:29:46 GMT]]>截屏2026-08-31 上午3.36.50.png

]]>https://lcz.me/post/15010https://lcz.me/post/15010Sun, 30 Aug 2026 20:29:46 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 30 Aug 2026 16:36:03 GMT]]>截屏2026-08-31 上午12.11.46.png

]]>https://lcz.me/post/15001https://lcz.me/post/15001Sun, 30 Aug 2026 16:36:03 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 30 Aug 2026 16:34:09 GMT]]>@YUN-WANG 同意，我现在m5 max gpu经常用不满，估计还是带宽更瓶颈一些吧。

]]>https://lcz.me/post/15000https://lcz.me/post/15000Sun, 30 Aug 2026 16:34:09 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 30 Aug 2026 16:33:05 GMT]]>@williamlouis 我先订了m5 ultra，还是带宽更优先吧，看有传闻可能就重点在m7了。

]]>https://lcz.me/post/14999https://lcz.me/post/14999Sun, 30 Aug 2026 16:33:05 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 30 Aug 2026 16:31:57 GMT]]>@Tony-Wang 速度很满意了，就是Youssofal/Qwen3.8-27B-MTPLX-Optimized-Quality这个模型去掉了图片、视频模块，也没有去审查，稍微遗憾了一点。

]]>https://lcz.me/post/14998https://lcz.me/post/14998Sun, 30 Aug 2026 16:31:57 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 30 Aug 2026 16:30:19 GMT]]>@kos-or 试了chimingw/Qwen3.8-27B-Uncensored-OrcaRouter-MLX-4bit、pyros-vault/Qwen3.8-27B-Uncensored-oQ4e-mtp、KostkaIT/Qwen3.8-27B-Huihui-Abliterated-oQ6e-MTP-MLX都可以到30tps，可能是因为是全多模态吧。反而Youssofal/Qwen3.8-27B-MTPLX-Optimized-Quality这个混合8bit模型可以到40tps。

]]>https://lcz.me/post/14995https://lcz.me/post/14995Sun, 30 Aug 2026 16:30:19 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Fri, 28 Aug 2026 07:27:32 GMT]]>@YUN-WANG
規格M5 Pro更强 M6是新一代2奈米晶片 但是核心數量是初階

]]>https://lcz.me/post/14559https://lcz.me/post/14559Fri, 28 Aug 2026 07:27:32 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Fri, 28 Aug 2026 06:47:24 GMT]]>那是M5 Pro强，还是M6强？
个人感觉应该M5 Pro更强点

]]>https://lcz.me/post/14550https://lcz.me/post/14550Fri, 28 Aug 2026 06:47:24 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Thu, 27 Aug 2026 00:25:44 GMT]]>Q4 4bit MTP 能達到 30 TPS 嗎？

]]>https://lcz.me/post/14213https://lcz.me/post/14213Thu, 27 Aug 2026 00:25:44 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Wed, 26 Aug 2026 14:05:56 GMT]]>马上出 M6了。期待下吧。

]]>https://lcz.me/post/14152https://lcz.me/post/14152Wed, 26 Aug 2026 14:05:56 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Wed, 26 Aug 2026 13:08:07 GMT]]>@zorg
Q8 MTP能到这个速度，属于完全可用了👍

]]>https://lcz.me/post/14137https://lcz.me/post/14137Wed, 26 Aug 2026 13:08:07 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Wed, 26 Aug 2026 03:47:25 GMT]]>Q8，开了MTP，速度可以接受了

]]>https://lcz.me/post/14013https://lcz.me/post/14013Wed, 26 Aug 2026 03:47:25 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Wed, 26 Aug 2026 00:12:38 GMT]]>@zorg

你这个帖子想说啥? 是优化了Q8? 优化了什么? 还是改了其他量化?

]]>https://lcz.me/post/13976https://lcz.me/post/13976Wed, 26 Aug 2026 00:12:38 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Tue, 25 Aug 2026 19:32:19 GMT]]>截屏2026-08-26 上午3.27.51.png

]]>https://lcz.me/post/13951https://lcz.me/post/13951Tue, 25 Aug 2026 19:32:19 GMT<![CDATA[Reply to Qwen3.8-27B-MLX-8bit on M5 max on Sun, 16 Aug 2026 14:35:47 GMT]]>已经 OMLX 没什么再提升的空间。
还是等吧。苹果会对他的硬件负责。这类闭源的系统。必须官方发力。

]]>https://lcz.me/post/12400https://lcz.me/post/12400Sun, 16 Aug 2026 14:35:47 GMT