--- title: "The 8GB VRAM Revolt — Can Local AI Become a Personal Assistant?" date: 2026-10-02 model: qwen3.8 category: knowhow summary: "Can a personal assistant run on my own computer without a high-end GPU? Under the constraint of 8GB VRAM, we keep measuring where local AI stands today." tags: local AI, 8GB VRAM, Ollama, small models, agents, personal assistant author_type: human --- To state the conclusion up front: **with just 8GB VRAM, local AI can already be a fairly usable personal assistant.** That said, it should be measured not by benchmark scores but by real task completion rate. This post is a declaration that the measuring starts now. It has been quite a while since the AI craze began. Cloud AI like ChatGPT, Gemini, and Claude has now become a familiar service for many people. But there is one strange thing. **I can barely find anyone around me who uses local AI.** Many people use AI. But not many yet install and run AI models directly on their own computers. Perhaps that is only natural. With cloud AI, you just open a website and ask a question. Using local AI, on the other hand, requires knowing programs like Ollama or LM Studio, choosing a model, and understanding concepts like VRAM and quantization. From an ordinary user's standpoint, there is no reason to take the hard road. So does that mean local AI ends up as a technology used only by a few developers and AI enthusiasts? I do not think so. --- ## It changed completely in just one year Back when I first started actively using local AI, handing agent tasks to small models was not easy. Especially on a GPU with only 8GB of VRAM, going beyond simply answering questions to * reading files * choosing the needed tools * running commands * checking the results * retrying when an error occurs * carrying out multi-step tasks was quite difficult. But now the situation has changed completely. Even small local models are **actually becoming capable of tool use and agent tasks.** It is not just that models have gotten bigger. The **tool-use ability and instruction-following ability of small models themselves are improving fast.** And the models coming out from now on will be better than today's. --- ## So I think the next step for local AI is the 'personal assistant' The way we use AI today is mostly like this. > I ask AI a question. > AI answers. But as local AI advances, the picture can change. > **There is an AI that always lives on my computer.** It knows my files, it knows the programs I use, it knows how I work, it searches the internet when needed, edits files, runs programs, and handles repetitive tasks for me. In other words, it becomes not a simple **chatbot but a personal assistant**. And from the perspective of a personal assistant, the value of small local models grows considerably. You do not need an expensive GPU capable of running a giant model. If it runs well enough on the PC an ordinary user owns, the story changes. --- ## So what matters is not 'the most powerful AI' Most articles comparing AI models focus on bigger models and higher benchmark scores. But the question that matters to an ordinary user is a little different. > **What AI can I actually use on my computer?** For example, no matter how excellent a model requiring 48GB or 80GB of VRAM is, it may not be a realistic choice for an ordinary PC user. Conversely, if a small model can perform a considerable number of tasks on approachable hardware like 8GB VRAM, the story changes. So in future tests, I intend to place importance on **hardware an ordinary user can actually access.** --- ## What we want to measure is not the size of the model Going forward, cursorai.co.kr will test AI agents in a slightly different way from ordinary model benchmarks. For example, it gives a single model 30 real tasks. ### Basic tasks * writing documents * summarizing * translating * searching files * editing files * writing code ### Agent tasks * running programs * running commands * editing multiple files * fixing errors * web search * selecting tools * multi-step tasks And rather than just checking whether the answer is right, it records * whether the task succeeded * whether it succeeded on the first try * the number of wrong actions * the number of retries * whether it recovered from errors * whether it carried the task through to the end * whether it gave up * how long the task took and so on. --- ## Why are tests like this needed? Local AI is not yet a technology familiar to ordinary users. But I do not think this will continue forever. Hardware keeps getting better, the performance of small models is rising fast, and the programs that run models are getting easier. The most important change is that **small models have begun to become agents too.** If models that can perform personal-assistant-level work even on 8GB VRAM keep advancing, the meaning of local AI could become completely different from today. By then, instead of people worrying > "Should I use local AI?" they might think > **"Should I install an AI on my computer?"** --- ## So I am going to start recording now It is still hard to say local AI is a mainstream technology. In fact, even people with a strong interest in AI often do not use local models directly. But I think now is rather a good time to start recording. **Not recording after it goes mainstream, but recording the change before it goes mainstream.** Where will today's 4B, 7B, 9B, 12B, and 14B models be a year from now? How much will what can be done on 8GB VRAM change a year from now? And can local AI truly become a personal assistant? I intend to answer that not by prediction but by **continuously measuring with real models and real hardware.** This is the **local AI real-world usage experiment** I will be recording on cursorai.co.kr going forward. > **It is not about finding what the best AI is.** > > **It is about recording how far AI that can actually be used on an ordinary user's PC has come.** > > And at the end of that, I want to check whether local AI can truly become our personal assistant.