I stopped using Claude to process sensitive files and switched to a local model instead

Published Sep 5, 2026, 2:00 PM EDT Nick Lewis is an editor at How-To Geek. He has been using computers for 20 years --- tinkering with everything from the UI to the Windows registry to device firmware. Before How-To Geek, he used Python and C++ as a freelance programmer. In college, Nick made extensive use of Fortran while pursuing a physics degree. Nick's love of tinkering with computers extends beyond work. He has been running video game servers from home for more than 10 years using Windows, Ubuntu, or Raspberry Pi OS. He also uses Proxmox to self-host a variety of services, including a Jellyfin Media Server, an Airsonic music server, a handful of game servers, NextCloud, and two Windows virtual machines. He enjoys DIY projects, especially if they involve technology. He regularly repairs and repurposes old computers and hardware for whatever new project is at hand. He has designed crossovers for homemade speakers all the way from the basic design to the PCB. Nick enjoys the outdoors. When he isn't working on a computer or DIY project, he is most likely to be found camping, backpacking, or canoeing. Every time you upload a file to a cloud AI service, you trust that they'll be responsible stewards of that information. When you're talking about medical data, financial records, personal details about your life, or any other sensitive information, that is a big ask. Fortunately, you don't need to rely on cloud AI services nearly as much these days if you own a moderate gaming PC. Local is better, even if you trust the company AI companies are an incredibly valuable target Credit: Nick Lews / How-To Geek Even if you fully trust that the big AI companies aren't going to do anything weird or nefarious with your data, they pose a different kind of risk entirely: they're an extremely lucrative target. The big AI companies have an extraordinary concentration of sensitive information about millions of people, including medical and financial data, personal thoughts, ideas, and more. If someone were to catastrophically compromise Anthropic or OpenAI, it would be one of the most significant data breaches in history. It isn't likely—these companies can afford to hire the best of the best in any field—but the risk is never zero. Furthermore, policy changes do happen. Terms of service and retention policies change. A promise made today isn't a guarantee of how your data will be handled in ten years down the road. On the other hand, a file that never leaves your machine has none of those risks. The only attack vector is your own computer, which already stores plenty of data in the first place. The model to use for local AI Qwen3.8-27B is practically a genius If you are looking for a starting point, Qwen3.8-27B is the model to use. It is recent (August 2026) and incredibly capable for its size. One of its biggest advantages over some older models is native image and video handling. You can give it scanned documents or photographs without needing a separate OCR program to translate for the model in advance. It can handle that natively. It also has a huge context window—theoretically up to about 262,000 tokens, though real-world performance falls off well before you hit that limit. Because it is available in a range of quantizations, you can get a pretty capable AI running on your PC while only using up 17GB of storage. Qwen 3.8-27B runs on most modern gaming PCs Your existing hardware may be enough Credit: Nick Lewis / How-To Geek If you have a common gaming build with 32 GB of RAM and a 12 or 16 GB GPU, you can run the 27B model at 4-bit. At that quantization, the model splits between VRAM and system RAM. This means generation happens at a pace roughly equal to how fast a person reads, not the near-instant speed of Claude. It is slower, but it is fully capable of contract review and statement analysis, and your data never leaves your PC. You could run the whole model on a 24GB GPU. If you can find a used RTX 3090, it is the most cost-effective way to run 27B parameter models locally today. Just make sure that the GPU you buy still has the VRAM. It is alarmingly common for people to strip them for VRAM and then dump them on Facebook Marketplace for $150–$300. My own PC has a RTX 5070 Ti with 16GB of VRAM and 32GB of system RAM, giving me a total of 40GB available for AI to use once you account for the operating system's overhead. Fortunately, that is plenty. I fired up Qwen 3.8-27B-q4 in Cherry Studio to put it through its paces. I provided a synthetic health record, a sample contract between a business and a vendor, and a divorce agreement, and then started quizzing it about what I found. My test revealed a few important things. The big one is that Qwen 3.8 27B-q4 is not especially prone to hallucinations. When I asked it about specific details in the files I'd provided to it, it didn't misquote anything. It was able to extract specific clauses, flag unusual values in medical data (which are flagged on the synthetic health record), and correctly tell me who was getting specific marital assets. It was a bit slow. The longest document was the contract, which was some 40 pages long. It spends a lot of time thinking between answers, but if you had some kind of bulk processing set up, that wouldn't be an issue at all. You could just leave it on while you make a cup of coffee. The only real limit is that thinking tokens share the context window with the document. That can lead to memory problems for very large files. Claude knows way more Local models are smart but ignorant Credit: Adam Davidson / How-To Geek The local model is excellent at reading your document, but cloud models still excel when it comes to knowledge. For example, my synthetic health record had an elevated AST—a potential sign of liver problems. Qwen 3.8-27B was perfectly capable of telling me the AST value and comparing it to historical trends, but what it couldn't do is meaningfully interpret it in the context of human health. "What does an elevated AST mean when ALP and ALT are normal"—a question you may be inclined to ask if you spend too much time scouring WebMD—could be reasonably answered by Claude but not by Qwen. Similarly, Claude seems to know a lot more about divorce laws in every jurisdiction than Qwen does. In other words, Qwen is actually excellent at extracting data and reasoning, but it doesn't know very much about the world. It is a bit like having a conversation with a 5-year-old with a 200 IQ and impeccable grammar. Keep your sensitive data local With the right hardware and the right local model, you can analyze all sorts of information without ever touching a cloud-based AI service. That'll probably improve over the next few years too. Frontier models have made exceptional progress in their capabilities over the last few years, and the same trend has held true for local models too. Early local AI models could scarcely put together a coherent sentence, much less accurately analyze something you provided. Today, there are at least a dozen models that will run on your gaming PC that would put the first public release of ChatGPT to shame.