You donโt need a graphics card or a cloud account to chat with an AI privately. This is a small, โuncensoredโ model that comes as one file with no signup โ download it, open it in a free app, and it runs right on an ordinary old laptop. ![]()
![]()
Start here โ pick your app
| Your situation | What to do |
|---|---|
| LM Studio โ paste the model link, click download, chat. | |
| GPT4All โ built-in model downloader, CPU-only. | |
Ollama โ install, then one ollama run line. |
|
| Jan โ auto-detects your hardware, no account. |
Then grab one model file below, or just paste its Hugging Face link straight into LM Studio.
Before you start: itโs not magic. 8 GB RAM is the real floor โ 4 GB laptops canโt run a model this size. The small โ3.8 GBโ file is a low-quality 2-bit version; the better Q4 is ~5.3 GB. It runs slowly on old CPUs (about 2โ12 words/sec), and โuncensoredโ means fewer refusals โ not smarter, and not always accurate.
What this model actually is
Losewโs Nemotron E4B Heretic is a free-to-download GGUF of NVIDIAโs Nemotron Bolt E4B โ a 4B-effective model built on Gemma 4 โ after the Heretic edit that reduces refusals. Itโs an unofficial derivative, free for personal and business use, and runs on plain llama.cpp.
| File | Size | Fits |
|---|---|---|
| IQ2_M (2-bit) | 3.79 GB | ~6โ8 GB RAM, lowest quality |
| IQ3_XXS (3-bit) | 3.98 GB | ~8 GB RAM |
| Q4_K_M (the good one) | 5.30 GB | ~8 GB RAM, best balance |
| Q8_0 | 7.97 GB | 12 GB+ RAM |
Get the file from the Files tab โ click the .gguf, then the download arrow. No account needed. Upstream: grumd0โs BF16 source.
Got an even older PC?
The Heretic edits exist on much smaller models too: Qwen 0.5B Heretic is a 398 MB file that runs under 1 GB RAM โ far less capable than the 4B, but it fits almost anything. Another option: tinyopsec 0.5B (Apache licence).
Find a model that fits your machine
Browse all Heretic models on Hugging Face โ about 6,700 to choose from; ggufmodels.com shows file sizes and memory estimates.
Not sure what will fit? CanRun (has a CPU-only mode), LocalLM Advisor (8 GB filter), and FitLLM all check for you.
Other small models worth a look: Qwen3-4B Heretic (~2.5 GB), Granite-3B Heretic (Apache 2.0), Nemotron Nano 4B.
LocalMode lets you try tiny models right in your browser; gguf-my-repo makes a model smaller; Unsloth Dynamic GGUFs keep more quality at the same size.
Make it a real chat app
Chaty gives a local model a normal chat window with your docs, voice, no account.
lilbee answers from your own files with citations.
Prefer one portable file? llamafile runs a model with a double-click (needs an AVX CPU).
Good to know
OpenMDW 1.1 (NVIDIAโs licence) allows personal and commercial use โ just keep the licence files; Gemmaโs base is Apache 2.0.
โUncensoredโ can drift quality and produce wrong or biased answers โ keep a stock model for serious work.
Large GGUF files arenโt virus-scanned by Hugging Face: download from known repos, check SHA256SUMS, and use a current llama.cpp.
Hugging Face blocked where you live? hf-mirror.com is the usual workaround.
Free to download and use, unofficial derivative โ not an NVIDIA or Google product. ![]()

!