Skip to the content.

在旧硬件上高效部署Gemma 4

该项目展示了如何在13年老的Xeon处理器上,不使用GPU的情况下,以每秒5个token的速度运行一个26B AI模型,展示了在旧硬件上高效部署AI的技术。 该项目因其高参与度(270分,175条评论)和在Hacker News上的讨论而具有重要意义,表明了人们对高效AI部署的浓厚兴趣。它解决了开发者在寻求成本效益AI部署时的实际问题,并通过SaaS或咨询服务具有潜在的盈利能力。 该项目使用tok

项目链接:https://www.neomindlabs.com/2026/06/08/running-gemma-4-26b-at-5-tokens-sec-on-a-13-year-old-xeon-with-no-gpu/ 作者:neomindryan 发布时间:2026-07-15T15:34:05Z 挖掘日期:2026-07-16 AI 评分:8.0/10 来源:hackernews 标签:AI, Efficiency, LegacyHardware, Inference, TokenOptimization

📌 项目详解

该项目展示了如何在13年老的Xeon处理器上,不使用GPU的情况下,以每秒5个token的速度运行一个26B AI模型,展示了在旧硬件上高效部署AI的技术。 该项目因其高参与度(270分,175条评论)和在Hacker News上的讨论而具有重要意义,表明了人们对高效AI部署的浓厚兴趣。它解决了开发者在寻求成本效益AI部署时的实际问题,并通过SaaS或咨询服务具有潜在的盈利能力。 该项目使用token优化技术来高效运行26B模型。它适用于希望在旧硬件上部署大型AI模型的开发者,而无需进行重大修改。

🌐 背景与生态

在旧硬件上运行大型AI模型一直是一个挑战,但token优化的进步使其成为可能。该项目基于这些进步提供了一个针对旧系统的实用解决方案。

💬 社区讨论

社区评论强调了在消费级硬件上运行大型模型的潜力,并讨论了本地与云端推理的成本影响。人们对效率的提升和在旧系统上运行强大模型的可能性感到兴奋。

🚀 应用前景

该项目在成本效益AI部署至关重要的行业(如医疗保健、教育和中小企业)中具有强大的应用前景。它可能导致开发能够普及高级AI能力的产品或服务。

🔧 技术栈

该项目利用了token优化技术,可能涉及像Python这样的编程语言,以及针对高效AI推理进行优化的框架和库。

🎯 上手难度

入门评级为进阶。前提条件包括13年老的Xeon处理器、Python以及熟悉token优化。基本步骤包括设置环境和运行示例推理脚本。

👥 目标用户

目标用户是希望在旧硬件上部署AI的开发者和企业。角色包括后端工程师、数据科学家和IT专业人员。

⚖️ 类似项目对比

竞品包括像’在MacBook Air上运行Qwen3.6-35B-A3B’和’Token优化用于AI推理’这样的项目。这些项目专注于高效运行大型模型,但在硬件要求和优化技术方面可能有所不同。

📚 参考链接

📄 查看原文内容 --- Top Comments --- [dwa3592]: I have a prediction. By the mid of 2027, we will have >200B MoE models running on basic consumer hardware. I am running Qwen3.6-35B-A3B locally on my 16GB mac with 7-9 tokens/second. Link - https://github.com/deepanwadhwa/samosa-chat This is a GPT4 level model running locally with a decent speed on a 16gb ram macbook air. [hagen8]: Some ppl don't like to hear it. But I would assume that token costs when using an inference provider are cheaper than electricity of using locally. If we just take into account output token generation for simplicity. With 5tps u get 18k tokens an hour. That would costs around 0.005USD from an inference provider. I estimate that the server consumes probably around 500W during inference. In Germany where 1kwh cost around 0.3USD, 18k tokens inferred locally would therefore cost 0.15USD whic... [Aurornis]: A dual Xeon of this era is probably pulling 300W or more when loaded. At national average electricity prices, that’s $1.35 per day. More during the summer if you have to cool the space. If you run it 24/7 and ignore prompt processing time (not a good assumption at all) it would get around 400,000 tokens in a day. That’s about $0.30 per million output tokens. Coincidentally, that’s the same price for this model on OpenRouter right now, but OpenRouter token gen will be 8X faster. There are... [hparadiz]: Here's my report running several different models on a dual Xeon with 256 GB of DDR4 and no GPU. https://gist.github.com/hparadiz/f3596d00a62d8ebb2dadcc46ee5... [throwaway2027]: That's quite slow I'm getting 8-12 t/s on a 13 year old CPU. (Speed varies by context size and other settings who knows) https://news.ycombinator.com/item?id=48354801