← 全部模型 Qwen · 近 7 天Qwen3.8 Flash Next0–100 分，越高代表社区使用体验越正面，不代表能力测试成绩。更新于 2026-09-05 16:01 UTC · 2026-08-29 – 2026-09-05 UTC社区口碑分61.1/100106 条评论 · 近 7 天 社区评论精选近 7 天的不同意见，摘录条数不代表真实好评比例。全部摘录正面负面1 条精选摘录 中性速度与延迟2026-09-05 “提示词处理在较低的PCIe带宽下变化很大，而令牌生成受的影响则小得多” Redditen机器翻译 展开原文Prompt prosessing changed a lot at lower pcie bandwidth while token generation was influenced a lot less正面推理2026-09-04 “这是一个前沿级模型” Redditen机器翻译 展开原文This is a frontier-class model中性本地部署2026-09-04 “Qwen3.8 flash next你可以用Q4甚至Q5，但你实际上需要64 GB系统内存，因为该模型中有51B n-gram。” Redditen机器翻译 展开原文Qwen3.8 flash next you can probably do a Q4 or maybe even Q5, but you'll actually need 64 GB system RAM because of the 51B n-gram in that model.正面推理2026-09-03 “它的推理能力很好，涵盖了我以前用前沿模型做的90%的事情。” Redditen机器翻译 展开原文It reasons well, it covers 90% of what frontier models used to do for me.正面推理2026-09-03 “目前是MOE的王者” Redditen机器翻译 展开原文king of MOEs right now负面通用文本2026-09-03 “Qwen3.8 flash next在许多推理引擎中的实现非常不稳定。再给它一些时间。” Redditen机器翻译 展开原文Qwen3.8 flash next's implementation is very unstable in many inference engine. Give it more time.正面通用文本2026-09-03 “从DS4F切换到它用于RAG摘要器，质量肯定提升了，没有乱码。vLLM on sparks，混合NVFP4+FP8 checkpoint。” Redditen机器翻译 展开原文Switched to it from DS4F for RAG summarizer, quality is definitely up, no garble. vLLM on sparks, hybrid NVFP4+FP8 checkpoint.负面推理2026-09-03 “qwen3.8-flash-next在xhigh档位下思考特别长，比glm-5.3-flash的max档位长太多了” 小红书zh 正面角色扮演/创作2026-09-02 “写创意故事方面做得非常出色” Redditen机器翻译 展开原文doing a really good job writing creative stories正面本地部署2026-09-01 “Qwen 3.8 Flash Next是朝着这个方向迈出的一大步” Redditen机器翻译 展开原文Qwen 3.8 Flash Next is a big step in this direction正面编程2026-09-01 “对它处理新颖提示的能力印象深刻” Redditen机器翻译 展开原文pretty impressed with what it can do with novel prompts负面通用文本2026-09-01 “使用nvfp4 radix量化后，质量相比deepseek有所下降” Redditen机器翻译 展开原文using nvfp4 radix quant the quality suffered compared to deepseek负面通用文本2026-08-31 “我一直在使用Qwen3.8 Flash Next，虽然对话水平不算惊艳，但我在自己的自定义测试环境中从未见过这种情况。” Redditen机器翻译 展开原文I have been working with Qwen3.8 Flash Next, and while the level of speech is not amazing, I haven't seen anything like this in my custom harness.正面本地部署2026-08-30 “Q8版本可以装在256GB系统中” Hacker Newsen机器翻译 展开原文fits on a 256GB system in Q8 近 7 天暂时没有符合此筛选条件的精选评论。查看更多摘录
中性速度与延迟2026-09-05 “提示词处理在较低的PCIe带宽下变化很大，而令牌生成受的影响则小得多” Redditen机器翻译 展开原文Prompt prosessing changed a lot at lower pcie bandwidth while token generation was influenced a lot less
中性本地部署2026-09-04 “Qwen3.8 flash next你可以用Q4甚至Q5，但你实际上需要64 GB系统内存，因为该模型中有51B n-gram。” Redditen机器翻译 展开原文Qwen3.8 flash next you can probably do a Q4 or maybe even Q5, but you'll actually need 64 GB system RAM because of the 51B n-gram in that model.
正面推理2026-09-03 “它的推理能力很好，涵盖了我以前用前沿模型做的90%的事情。” Redditen机器翻译 展开原文It reasons well, it covers 90% of what frontier models used to do for me.
负面通用文本2026-09-03 “Qwen3.8 flash next在许多推理引擎中的实现非常不稳定。再给它一些时间。” Redditen机器翻译 展开原文Qwen3.8 flash next's implementation is very unstable in many inference engine. Give it more time.
正面通用文本2026-09-03 “从DS4F切换到它用于RAG摘要器，质量肯定提升了，没有乱码。vLLM on sparks，混合NVFP4+FP8 checkpoint。” Redditen机器翻译 展开原文Switched to it from DS4F for RAG summarizer, quality is definitely up, no garble. vLLM on sparks, hybrid NVFP4+FP8 checkpoint.
正面角色扮演/创作2026-09-02 “写创意故事方面做得非常出色” Redditen机器翻译 展开原文doing a really good job writing creative stories
正面本地部署2026-09-01 “Qwen 3.8 Flash Next是朝着这个方向迈出的一大步” Redditen机器翻译 展开原文Qwen 3.8 Flash Next is a big step in this direction
正面编程2026-09-01 “对它处理新颖提示的能力印象深刻” Redditen机器翻译 展开原文pretty impressed with what it can do with novel prompts
负面通用文本2026-09-01 “使用nvfp4 radix量化后，质量相比deepseek有所下降” Redditen机器翻译 展开原文using nvfp4 radix quant the quality suffered compared to deepseek
负面通用文本2026-08-31 “我一直在使用Qwen3.8 Flash Next，虽然对话水平不算惊艳，但我在自己的自定义测试环境中从未见过这种情况。” Redditen机器翻译 展开原文I have been working with Qwen3.8 Flash Next, and while the level of speech is not amazing, I haven't seen anything like this in my custom harness.
社区评论
精选近 7 天的不同意见，摘录条数不代表真实好评比例。
1 条精选摘录
展开原文
Prompt prosessing changed a lot at lower pcie bandwidth while token generation was influenced a lot less
展开原文
This is a frontier-class model
展开原文
Qwen3.8 flash next you can probably do a Q4 or maybe even Q5, but you'll actually need 64 GB system RAM because of the 51B n-gram in that model.
展开原文
It reasons well, it covers 90% of what frontier models used to do for me.
展开原文
king of MOEs right now
展开原文
Qwen3.8 flash next's implementation is very unstable in many inference engine. Give it more time.
展开原文
Switched to it from DS4F for RAG summarizer, quality is definitely up, no garble. vLLM on sparks, hybrid NVFP4+FP8 checkpoint.
展开原文
doing a really good job writing creative stories
展开原文
Qwen 3.8 Flash Next is a big step in this direction
展开原文
pretty impressed with what it can do with novel prompts
展开原文
using nvfp4 radix quant the quality suffered compared to deepseek
展开原文
I have been working with Qwen3.8 Flash Next, and while the level of speech is not amazing, I haven't seen anything like this in my custom harness.
展开原文
fits on a 256GB system in Q8
近 7 天暂时没有符合此筛选条件的精选评论。