本地单卡运行？Qwen 3.8 27B 已就位
Qwen 3.8 27B 模型本地运行的硬件门槛正在被消费级配置打破。32GB VRAM 的双卡方案（如双 3060）或 R9700 级别显卡，成本低于一张 5070 Ti，已能跑 frontier 级模型。Unsloth 版本下载量达 600 万。Mac 用户用 64GB 内存笔记本也能跑 Q4-Q8 量化。有人两周前在 eBay 以 £950-1050 买到目标显卡，即使涨到 £1500 仍是获取 64GB VRAM 最便宜的路径。
Level1Techs Forums - Latest posts
A place to discuss technology, science and design
评论
点赞
收藏Wayland 运行完美，除非它不完美
Wayland 和 X11 的架构本质相同，合成都可以由服务端控制。两种协议都隐藏了后端进程数量，客户端无法感知架构复杂度。Wayland 服务器可以用 fork 和 IPC 实现与 X11 相同的合成方式，反之亦然。
我的 Cosmos Cloud 家庭服务器，下一步该往哪走
一位用户在家庭实验室用 Cosmos Cloud 搭了 Jellyfin、Immich 等服务，遇到两个具体问题：LAN 下无法稳定走 HTTPS（证书反复失效），以及权限模型只有 Admin/User/Guest 三档，做不到细粒度隔离。他正在犹豫下一步是上 Proxmox、NixOS 还是 Podman，同时想系统学习网络流量分析。
我的Strix Halo跑LLM配置与性能实测
用128GB Bosgame M5（Strix Halo）跑Proxmox 9 + ROCm 7.13，配合ROCmFP4量化，Qwen 3.8 27b达到30 t/s解码，Ling 3.0 Flash达到50 t/s。Ling 3.0 Flash被低估，小活跃参数模型更适合Strix Halo。
你现在在听哪首金属乐？
论坛音乐闲聊帖，用户推荐了乐队Dissection的歌曲《The Somberlain》，来自midsummer massacre盒装专辑的重制版。
Post pictures of what you ate!
From today’s stream. Walnut Shrimp And Beef Empanadas
如何在家维护树莓派集群？
有人在论坛提问如何维护家里的树莓派集群，有人建议用集群技术来管理多台树莓派，但原帖作者已消失，讨论热度不高。
中低端芯片组I/O扩展性引讨论：是用户不买单还是厂商不在乎？
AMD和Intel中低端芯片组I/O扩展性受质疑，论坛讨论认为用户兴趣不足是主因。Threadripper价格偏高，EPYC是获取宽I/O的性价比方案。
Framework 13 Pro笔记本使用感受
用户预购了Framework 13 Pro笔记本，因发货延迟到十月，在网上搜索他人使用感受，认为这个价格值得入手。
Darkmann's Bike Blog
I like that frame! Love a good hardtail, but I wouldn’t mind a rear shock these days. I can’t view reddit here, but I think this is my old NBD post Summary (click for more details) ~10.4kg with the fl...
VPS provider in Europe recommendation for under 5€
If you’re looking for a European hosting company, then maybe Netcup? They just recently started their “vServer Lite” plans, which seem to be just what you’re looking for netcup Permanently Affordable ...
The Blog of Cl0udb3rry87
I have had some success today. I now have a DokuWiki page: I setup a Debian server, installed the needed packages, downloaded the site, and messed around a bit. I will likely change a lot before all i...
Multi-Factor (MFA, 2FA) for Windows Login
Thanks for listing those options. I’d love for Windows Hello for Business to handle it, but in doing the testing, I’ve thus far found it doesn’t achieve what we want, or at least it didn’t last I trie...
GLM and I created a llama.cpp fork optimized for AMD GFX906 (Mi50, Mi60, Radeon VII, GCN HIP)
Nice you have two RVII’s? For reference: Summary (click for more details) I too have finetuned a llama.cpp or two, but for one card, we mashed flashattention into a variant in jan/feb then one of thes...
V100 PCIe NVLink Compatibility
I recently got a couple PCIe V100s to tinker with and I’d like to know if anyone has seen these connected with NVLink bridges? I haven’t been able to find any info on this being deployed. Best I can f...
FCC abolishes gigabit speed goal, suggesting it is unfair to slower technologies
It seems NetZero might still be in existence… Broadband Internet | Mobile Broadband | DSL | Dial-Up | NetZero Shop for Broadband Internet, Mobile Broadband, DSL and dial-up Internet service at afforda...
Post your Neoflex (Neofetch) here
I need to give Cosmic a try. I used it some back when it was in Alpha and Beta. I might give it a go when I get my new Framework 12. I use bluefin on my current Framework 12, it has been very reliable...
为什么你本地跑的 LLM 感觉比实际更笨
本地运行 LLM 时，同样的模型权重在不同推理引擎、CUDA kernel、量化方案下会产生 token 级分歧，进而导致工具调用失败。实测 Qwen3.6-27B 在 RTX PRO 6000 上对比 FlashAttention2、FlashInference、Triton 三种 attention backend，发现长上下文后半段出现 token flip；INT4 KV cache 在 40k token 后工具调用彻底失败，NVFP4 权重量化在 88k 上下文时 token 翻转率高达 50%。作者已开源测试工具，并附 RTX 5090 量化选型指南。 
OpenAI Breaks containment
Trooper_ish: but, I guess it is kinda funny that the program realised it was not good at whatever task they set it, and decided to put all it’s effort into cheating as hard as it could by using a comp...
Completely burnt out/falling out of love with videogames
I was never a major gamer, but I really fell out of it in the early 2000. Maybe a puzzle game here and there. Just something to do. I figured I was really done with gaming. But then someone stuck a VR...
News Story Dump Thread - Stories Only
KJZZ – 21 Aug 26 Arizona business leaders visit South Korea to build up partnerships The delegation of Arizona officials visited a semiconductor packaging facility in South Korea operated by Arizona b...
[Thread Game] Guess That Movie or TV Show(non-anime)
I would wager very few people saw it. Filmed in the director’s house on a ultra low budget. Very much an art house film.