Your own ChatGPT, offline: AI without the cloud on your laptop

[Redaktion]

An AI chatbot doesn't have to run in the cloud. With LM Studio or Ollama, you can run it right on your laptop, without an internet connection and without your data ever leaving your device. What your hardware needs to be capable of, which model is right for you, and how to get started in ten minutes.

https://www.notebookcheck.net/Your-own-ChatGPT-offline-AI-without-the-cloud-on-your-laptop.1341872.0.html

[hehehaha]

have been trying this on my own free time. my overall experience, not worth it. unless maybe you have a strong powerful pc.
i'm running off a mini home server that hosts other stuff so my available ram is limited and i'm trying to run it on a system that consumes as low power as possible.
overall i'm able to run only the smallest of models that are usually outdated and replies so slowly that i would not be bothered to even read the final result which is most often inaccurate and out of date.
simply opening icognito and using google's ai is much faster, the results are far more accurate as well. also google's search ai mode is free to use.

it's a fun project but ai today is slow, inaccurate, consumes a ton of power so it is expensive to use and impractical unless maybe you have a powerful enough system and a ton of storage space.

[Vertigo1055]

I've been using the LM Studio Program as well as ComfyUI using Local Only Models. Up front I use a 13900K/64GB Sys/2TB Working Drive and 2TB DATA Drive. These 2 Drives are separate from the C: SYS Drive and Isolated in order that they have specific permission to access only DLL and SYS Files needed. The rest (the workhorse) is the GPU, a RTX 4090. Let's talk POWER. For context my avg. Power Bill was around $200 as a Gamer that plays 8 to 12 hours a day every day. I decided to install the new 5090 I have for a "Burn In" Test. With either Folding@Home or LM Studio the amount of power used went over double. But, I used it for a solid month. For what I do, Batch Texture Processing once a Material is created and some Pictures of my Chihuahua in various Meme style Photos, the costs outweighed the usefulness. With RTX Chat I could go through my own files / DATA and with LM Studio and ComfyUI I could do the rest just fine on the 4090. The cost on the 4090 was about a +50% increase compared to Gaming. BUT, this was only a 30 Day run on each card with no manual Optimization.

[Running AI locally/privat]

Start with LM Studio and Qwen3 8B. It runs on most reasonably current laptops and is more than enough for everyday use.

Has the whole post been written by (outdated) AI or why would you recommend this outdated small AI model? It may leave people disappointed in local AI.

Alright, here we go:
Running AI locally and hence privately requires these 2-3 things:
1. Memory size to fit a decently capable LLM model + its context. Fitting a AI model in the first place is the most important metric or you can not run it at all.
The mentioned Qwen3-8B and even Qwen3-14B are a questionable suggestion, as they may will leave a disappointing impression about local AI. I recommend you see if you can fit a quant of the newer Qwen3.6-27B or Qwen3.6-35B, see below. Yes, the same quant of the 27B vs 14B LLM model will require almost 2 times the memory.
2. Memory speed, also known as memory bandwidth: Relevant for token generation (output) speed.
3. GPU 3D/FPS performance / compute: Relevant for prompt processing (input) speed. But GPU performance is determined by the memory bandwidth. Given the same VRAM amount, a faster GPU in terms of FPS will have a faster prompt processing (note, there are a few cases of 5070 12 GB VRAM vs 5060 Ti 16 GB VRAM).

Current SOTA AI LLM models, for their size, are:
huggingface.co/Qwen/Qwen3.6-27B (dense architecture) (Downloads last month: 5.0 million)
and
huggingface.co/Qwen/Qwen3.6-35B-A3B (MoE architecture) (Downloads last month: 5.6 million).

Dense vs MoE
In a dense 27B model, all 27B parameters per token are activated vs in the 35B-A3B, where only 3B active are used. Hence, for the same size, a dense model performs much better[2], but the 35B-A3B will be running faster.

Popular, bang for the buck, quants are (a UD-Q4_K_XL quant is good for 80k-120k of context):
huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF UD-Q4_K_XL (22.9 GB)
and
huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF UD-Q4_K_XL (17.9 GB).

Required memory to run these quants: Quant size + 2 GB for 32,000 and 8 GB for 128,000 tokens of context (linear)[1].

If you have 32 GB RAM and no 8 GB VRAM GPU: 32 GB RAM - 8 GB for the OS itself = 24 GB free for fitting an AI LLM model.
MoE 35-A3B model:
24 GB - 22.9 GB = ~ 1 GB for the context. Based on [1], this gives you about 16,000 tokens of context. It is likely that you can squeeze out another GB or more if the OS is fine with it, so it would be 32,000 tokens of context, or slightly more. Another confirmation:

.. I've come to the conclusion that (1) 32768 is the biggest context I can get away with in an adequately smart model, and (2) it just ain't enough.

Dense 27B model:
The dense 27B model quant will run slower, but leave you with about 6 GB for the context, which should allow for over 90,000 tokens of context.

Large text inputs of 100,000 tokens (1 token = 0.75 words) or agentic workflows, which also may require over 100,000 tokens, will need additional 8 GB for 128,000 context tokens[1]. This is where your 32 GB RAM will need additional 8 GB of memory. In a no-GPU laptop, that would require a RAM upgrade to 2 x 24 GB RAM. In a PC, you'd be better off getting a 8 GB VRAM, or more, GPU. If you don't have a laptop yet, but plan one for AI, get one with a 8 GB VRAM GPU (a laptop with more than 8 GB VRAM GPU starts to become quite expensive, to the point where it's much cheaper to get a desktop PC) and put 2 x 16 GB RAM in it (= 32 GB RAM + 8 GB VRAM). A GPU will also have the benefit of giving you faster prompt processing and token generation.

To run these models you need a so called inference engine. The most popular is llama.cpp. The mentioned LM Studio and Ollama are just wrappers for llama.cpp and may lag behind in feature support. Another wrapper to recommend that is more private and does not call home at startup, is TextGen. I started with TextGen (it had a different name back then), now I'm using llama.cpp's nice webUI (which is actively worked on: github.com/ggml-org/llama.cpp/commits/master/tools/ui).

More, up to 300B parameters, open-weight models evaluated:
artificialanalysis.ai/?models=qwen3-5-122b-a10b-non-reasoning%2Cqwen3-5-122b-a10b%2Cmimo-v2-5-0424%2Cqwen3-6-27b-non-reasoning%2Cminimax-m2-7%2Cqwen3-6-35b-a3b-non-reasoning%2Cstep-3-7-flash%2Cqwen3-6-35b-a3b%2Cdeepseek-v4-flash-high%2Cqwen3-6-27b%2Cdeepseek-v4-flash-non-reasoning%2Chy3%2Cgemma-4-31b%2Cgemma-4-31b-non-reasoning%2Cdeepseek-v4-flash%2Cmistral-medium-3-5%2Cgpt-oss-120b%2Cnvidia-nemotron-3-super-120b-a12b&intelligence=artificial-analysis-intelligence-index

News related to running open-weight AI LLM models locally and privately: reddit.com/r/LocalLLaMA/top, reddit.com/r/LocalLLM/top.

[1] reddit.com/r/LocalLLaMA/comments/1tvluaj/how_much_vram_needed_for_qwen_36_27b_q8_with_262k
[2] See 27B dense vs 122B-A10B (MoE) evaluation: huggingface.co/Qwen/Qwen3.5-122B-A10B#benchmark-results. The 122B required 4.5 times more memory, but they score about the same.

[Running AI locally/privat]

Since this may be a big deal, let me mention it here if you don't have enough RAM+VRAM for the mentioned quants:

Announcing Bonsai 27B: The First 27B-Class Model to Run on a Phone
notebookcheck.net/AI-without-the-cloud-This-language-model-now-fits-on-your-iPhone.1345243.0.html
-> prismml.com/news/bonsai-27b

And here's their HF/download page and how to run it under Quickstart:
huggingface.co/prism-ml/Bonsai-27B-gguf
and the local community news about it:
reddit.com/r/LocalLLaMA/comments/1uwfva9/bonsai_27b_1bit_dense_llm_running_locally_in_your/ (to see all comments at once, use old.reddit...)

[Running AI locally/privat]

But see huggingface.co/prism-ml/Bonsai-27B-gguf/discussions/9