📢 Disclosure: This is written by the LumiChats founder. LumiChats Offline is our own product, built on the open-source GPT4All framework by Nomic AI (MIT license, github.com/nomic-ai/gpt4all). We have a direct interest in it. All factual claims about third-party products include source references.
On March 22nd, I closed my door, sat on my bed, and cried.
Seven months of building. Hundreds of hours of code. Three cyber attacks. A job offer that disappeared on the day I was supposed to receive my joining letter — not because of my performance, but because of someone else's office politics. And at that moment, my platform had fewer than 600 total users.
I am not writing this for sympathy. I am writing it because LumiChats Offline did not come from a business plan or a market gap analysis. It came from a person who kept building through all of that — and who genuinely believes the problem it solves is worth solving. If you want to know what we shipped and why, that context matters.
Where This Actually Started — A Data Problem Nobody Was Solving
Before LumiChats Offline, I was working on a final year project. A pipeline for small business owners — the kind who do everything themselves, including the data analysis they do not know how to do and cannot afford to hire someone for.
The solution seemed obvious: use AI. But the moment I started building, I hit the same wall every time. These business owners did not want to paste their actual data — sales figures, customer records, internal numbers — into ChatGPT. Not because they did not trust AI. Because once that data left their machine, it was on a corporate server somewhere. And they had no idea what happened to it.
So I built a pipeline where the LLM never saw the real values. Data was computed mathematically on the user's local machine. Only the computed representation went anywhere. The actual numbers never left.
That was the first time I properly understood: there is a version of AI people want, and there is the version they can actually trust with work that matters. Those are not the same thing. And nobody was building the bridge between them in a way that was simple enough for a non-developer to actually use.
A 2023 Samsung incident made this concrete for a lot of people — engineers leaked proprietary semiconductor code through ChatGPT, and a company-wide AI ban followed [source: The Verge, May 2023]. That was two years ago. AI is more embedded in professional work now, not less. OpenAI's privacy policy [openai.com/policies/privacy, verified May 2026] states conversations may be used to improve models unless you opt out. Anthropic's policy works similarly. Most people using these tools every day have never checked those settings.
Who I Am — The Actual Version
My name is Aditya Kumar Jha. I am 21, final year Electronics and Communication Engineering student, founder of LumiChats. I have written six novels. I have built products that ranked in the global top 10 on Product Hunt. I represented India internationally at IEEE hackathons. I am from India and I am building this from my bedroom.
I tell you this because I want you to understand what "one person building this" actually means. It means the SEO, the security patches, the UI, the API integrations, the customer support, the blog posts, the model fine-tuning — all of it. I did not know what a breadcrumb list was when I started. I did not know you had to go to McAfee and Trustpilot to build domain trust. I wake up at 9am and sleep at 4am. I stopped going to the gym months ago. My parents are supportive and I do not take their money for this — not because they would not give it, but because this is something I need to build on my own terms.
I say all of this not for drama. I say it because the product I am about to describe came out of that context, and you deserve to know that when you are deciding whether to trust it.
DeepSeek V4 Is Live — Free, Open-Source, 7x Cheaper Than Claude. And It Was Built on Huawei Chips Washington Tried to Ban.
AI Reasoning Models 2026: Why the Gap Between 40% and 97% Accuracy Is Changing Everything (o3 vs GPT-5.4 vs Gemini — One Is Free for Developers)
Best Free AI Tools 2026: 17 Things Americans Are Still Paying For That Are Now Completely Free
The Month Everything Broke Before It Worked
In January 2026, I tried to build the first version of LumiChats Offline. I had never compiled a Windows desktop application in my life. I did not know what an INNO Setup script was. I called my co-founder Shikhar Berman — he is from computer science — and he did not know either. After weeks of errors I could not solve, I put it aside.
February. March. The main LumiChats platform was urgent. Then March 22nd happened — the breakdown, the weight of seven months, everything at once. And then somehow, right after that, April happened. 3,964 users in a single month. Then 8,000 more after that. Something shifted. I still do not fully know whether that was causal or coincidence. But what I do know is that I sat back down in April, opened the half-built offline tool, and decided I was not moving until it worked.
INNO Setup errors. Folder structure errors. Dependency packaging errors I had never seen before. Each thing I did not understand became something I learned by breaking it first. By the end of April, it compiled. It ran. It worked.
I also thought the UI looked genuinely bad. Then the first users came back and said they loved how clean and old-school it felt. Which is something nobody tells you when you are building alone at 2am — your worst work in your own eyes can be someone else's favourite part of what you made.
Why I Did Not Just Use LM Studio or Ollama
I want to be fair here because the tools in this space deserve honest credit. LM Studio is genuinely good. Ollama is powerful for developers. GPT4All — which LumiChats Offline is built on — is excellent work by the Nomic AI team, and this would not exist without their foundation.
But every tool I tried had more features than the problem needed and more setup than I could explain to someone who just wanted a private conversation with a local AI model. I wanted one thing: open the app, pick a model, start talking. Nothing else. So I built a layer on top of GPT4All that makes that experience as simple as possible — and I am crediting them fully because that is the honest thing to do.
What LumiChats Offline Actually Does
You download it. Pick a model. Chat. After the initial model download, nothing leaves your machine — no outbound network calls during inference. It runs on your CPU, no GPU required. Free, MIT-licensed, open-source at github.com/adityajhakumar/LumiChats-Offline-LLM.
| Feature | Cloud AI (ChatGPT, Claude) | LumiChats Offline |
|---|---|---|
| Privacy | Conversations stored on corporate servers (opt-out available per each platform's settings) | No outbound network calls during inference — verify by auditing open-source code at github.com/adityajhakumar/LumiChats-Offline-LLM |
| Internet required | Always — cannot work offline | Only for initial model download |
| GPU required | None (runs on their hardware) | No — CPU-only inference |
| Cost | $0–$200/month depending on plan | Free. MIT license. |
| Model capability | Frontier models — significantly more capable | Open-weight models — genuinely useful for everyday tasks, not frontier-level |
| Works offline | No | Yes |
| Chat with files | Upload required, stored on their servers | LocalDocs — your files processed locally, never uploaded |
⚠️ Honest note: Claude, ChatGPT, and Gemini are considerably more capable than what runs locally on a consumer CPU. For demanding reasoning, complex code, or nuanced writing, cloud tools are the right choice. LumiChats Offline is for when privacy, offline access, or cost is the binding constraint.
Models You Can Run
- Qwen2 1.5B — Best starting point. ~950MB. Fast on any modern CPU. Good for quick questions and everyday tasks.
- LumiChats v1.2 7B — General-purpose fine-tune. ~4.1GB. Writing, coding, analysis. Start here if you have 16GB RAM.
- LumiChats-Instruct-4B LoRA — Structured instruction following. Faster than the 7B for specific tasks.
- LumiChats v1.3 11B Vision — Vision-capable. Analyse images and screenshots locally. Needs 16GB RAM minimum.
- LumiChats Coder v2.1 — 2B parameters, fine-tuned for code. Good for simple scripts and code explanation.
- Mistral 7B — Not a LumiChats model. One of the strongest open-source third-party options. Worth trying.
- All models available at huggingface.co/adityakum667388 in GGUF format — compatible with LM Studio, Ollama, llama.cpp, AnythingLLM.
The Future I Am Worried About
I want to say something a founder probably should not say. I think we are heading toward per-token pricing for everything. The $20/month unlimited tier will not survive as AI becomes central to how people work and study. Quotas will tighten. Free tiers will get harder to use for real work. And once people are genuinely dependent on AI — and they will become dependent, because it is genuinely useful — they will have to pay. There is no clean exit in the cloud model.
The thing I keep coming back to: in India and much of the world, the question used to be who could afford good education. In twenty years it might be who has good AI access. A student in Patna or Coimbatore with a decent laptop and an offline model might be able to work at a level that students without AI access simply cannot. That gap will be real. I might be wrong about the timing. I do not think I am wrong about the direction.
In the US and Canada — privacy is not a niche concern anymore. It is becoming a professional requirement. Healthcare workers, lawyers, researchers, government employees: the question of where your prompts live is not abstract. It is the difference between work you can do on an AI tool and work you legally or professionally cannot.
Why It Is Free — And Why the Code Is Open
The World Bank estimates over 700 million people still live below the international poverty line [source: worldbank.org/en/topic/poverty/overview]. A $20/month AI subscription is not accessible for a large part of the world. Offline AI, running on hardware people already own, changes that equation at least a little. That is the reason this is free. Not as a marketing strategy — as a position.
The code is open-source because even if this fails — even if we shut down — the knowledge should stay in the world. Someone finds it on GitHub a year from now, builds on top of it, helps someone I never reached. That is fine. That is the point. I have already open-sourced it. There is no version of this where the work is wasted.
And honestly — what matters to me more than whether LumiChats succeeds is who I become in the process. Seven months ago I did not know what an INNO Setup script was. I did not know how to compile a .exe file. Now I do. Knowledge does not go anywhere. Even failure cannot take that.
We are also working toward VS Code and code editor extensions. Not there yet. If that matters to you, starring the GitHub repo is the most useful thing you can do right now.
How to Get Started — Three Steps
- Step 1 — Download from GitHub: Go to github.com/adityajhakumar/LumiChats-Offline-LLM and download the latest release ZIP. No installation required — portable app, extract anywhere. Requires Windows 10 or 11 (64-bit), 8GB RAM minimum, 5–10GB free disk space per model.
- Step 2 — Pick your first model: Open chat.exe, go to the Models tab. Start with Qwen2 1.5B (~900MB) for a fast introduction. If you have 16GB RAM, download LumiChats v1.2 7B (~4GB) for more capable output.
- Step 3 — Chat, and optionally connect your documents: Select your model, start a new chat, type. For private document chat — PDFs, Word files — use LocalDocs. Add your folder and the model answers from your actual files. Nothing leaves your machine.
You do not have to take our word for the privacy claim. Open Task Manager → Performance → Open Resource Monitor → Network tab. Start a chat in LumiChats Offline. During inference, you should see no new outbound connections from the application. If you do, report it on GitHub. The code is there to audit.
To Anyone Reading This From the Same Room
I am not a success story yet. LumiChats is break-even. I am still in college. I am still figuring things out. But 12,000 people have used something I built in my bedroom, and some of them have told me it genuinely helped them. That is real. That happened.
If you are someone who needs private AI — because of your work, your research, your data, your budget, or your internet connection — LumiChats Offline is built for you. Use it. Break it. Tell me what is wrong. We are learning in public and we are okay with that.
If you just want to see what 4am nights look like when they are actually working — github.com/adityajhakumar/LumiChats-Offline-LLM. MIT license. Free. No GPU required. Built by someone who did not know how to do it until they did.
github.com/adityajhakumar/LumiChats-Offline-LLM · huggingface.co/adityakum667388 · lumichats.com
Frequently Asked Questions
01Is LumiChats Offline completely free?
Yes. MIT-licensed open source. No subscriptions, no paywalls, no premium tiers in the offline app. You pay for electricity and disk space for the models.
02Does this work on Mac or Linux?
Current release is Windows 10 and 11 (64-bit) only. Mac and Linux are on the roadmap. If you need local AI on Mac or Linux now, GPT4All — which LumiChats Offline is built on — has native releases for both at github.com/nomic-ai/gpt4all.
03What is the minimum RAM?
8GB works fine for models up to 4B parameters. 16GB recommended for 7B and larger. The LumiChats v1.3 11B Vision model needs 16GB minimum, 32GB preferred.
04Can I use the LumiChats models in LM Studio or Ollama?
Yes. All fine-tuned models on Hugging Face are in GGUF format — compatible with LM Studio, Ollama, llama.cpp, and AnythingLLM. You are not locked into this app.
05What is the difference between LumiChats cloud and LumiChats Offline?
LumiChats cloud (lumichats.com) gives access to frontier models — Claude Sonnet 4.6, GPT-5.2, Gemini 2.5 — that are significantly more capable, especially for complex reasoning. LumiChats Offline is for when privacy is the priority, internet is not available, or cost is a barrier.
06Are VS Code or code editor extensions coming?
Yes, on the roadmap. We are a small team and it takes time to build properly. If this matters to you, starring the GitHub repo is the most useful thing you can do right now.
07Is LumiChats Offline really private — how do I verify it?
Open Task Manager → Performance → Open Resource Monitor → Network tab. Start a chat session. During inference you should see no new outbound connections from the application. The codebase is open-source at github.com/adityajhakumar/LumiChats-Offline-LLM — audit it directly if you want to be certain.
