Testing 3 Local LLMs on a CPU-Only Laptop — Here's What Actually Happened

一位开发者用一台没有独立显卡的普通笔记本，实测了三个开源大模型，结果发现本地AI跑代码生成任务并非不可能，但体验相当“酸爽”。最吸引人的地方在于，这打破了“没有昂贵GPU就玩不转AI”的固有印象，证明了40GB内存的CPU设备也能运行14B到26B参数的大模型。简单来说，这项测试就像让三个不同体量的“实习生”在无网环境下独立完成一个贪吃蛇小游戏。其中Qwen 3 14B表现中规中矩，但生成一次要等9分钟，修改一个小问题还得再来一轮；GPT-OSS 20B和Gemma 4 26B则各有千秋，但都存在生成速度慢、代码需手动调整的通病。整个过程完全离线，不花一分钱API费用。对普通人而言，这意味着未来在没网的飞机上或偏僻地区，也能用笔记本本地跑AI助手处理简单编程任务，隐私性也更好。不过目前来看，它更像是技术爱好者的玩具，距离替代云端AI还有很长的路要走。普通用户想流畅使用，依然得依赖联网服务。

As developers, we keep hearing that powerful LLMs require expensive GPUs. I wanted to find out how far local AI could actually go on a modest machine — no cloud, no cost, no GPU. So I ran a practical experiment using LM Studio on a 40 GB RAM laptop without a dedicated GPU, pitting three open-weight models against each other on a real coding task. <h2> Test Environment </h2> Hardware <ul> <li>Standard laptop</li> <li>40 GB RAM</li> <li>No dedicated GPU (CPU inference only)</li> </ul> Software <ul> <li>LM Studio (context length: 8192 tokens)</li> <li>VS Code</li> <li>100% local inference — no API calls, no internet required</li> </ul> Models tested <ul> <li>Qwen 3 14B</li> <li>GPT-OSS 20B</li> <li>Gemma 4 26B A4B QAT</li> </ul> <h2> The Experiment </h2> Each model received the exact same prompt: generate a complete Snake Game using only HTML, CSS, and JavaScript — no frameworks, no external libraries. <a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmuqq2at61higzaiiqc48.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmuqq2at61higzaiiqc48.png" alt=" " width="719" height="419"></a> I evaluated each result on: <ul> <li>⏱ Generation time</li> <li>🪙 Token usage</li> <li>🎨 Quality of the generated UI</li> <li>🔧 Whether manual corrections were needed</li> <li>🧑‍💻 Overall developer experience</li> </ul> <h2> Results </h2> <h3> 🟡 Qwen 3 14B </h3> Generation took approximately 9 minutes and produced 3 separate files (HTML, CSS, JS). After testing, the snake moved too fast. I had to send a follow-up prompt, which triggered another full 9-minute generation cycle — and only around five lines of code actually changed. <a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2ptagum0it56snlqwxaz.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2ptagum0it56snlqwxaz.png" alt=" " width="800" height="426"></a> <a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fc0s1oq5klkoyl6ct5ak4.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fc0s1oq5klkoyl6ct5ak4.png" alt=" " width="164" height="168"></a> Pros <ul> <li>Functional game on the second attempt</li> <li>Game instructions included</li> <li>Low token usage (~651)</li> </ul> Cons <ul> <li>Slow generation</li> <li>Required a second iteration to fix gameplay speed</li> <li>Basic UI</li> </ul> <h3> 🟢 GPT-OSS 20B </h3> This was the biggest surprise of the experiment. Generation completed in approximately 3 minutes, producing a single self-contained HTML file with all HTML, CSS, and JS included. Token usage clocked in at ~51 — significantly lower than the other models (worth noting: token counting may differ across models, so treat this as directional rather t