Why can I run Llama on my laptop but not GPT?
Open weights vs closed API models, and model licenses · Beginner · 5 min · llm, training · after Model size, free vs paid tiers, Quantization (bits per weight), Knowledge distillation (teacher and student models)
A model is a file of numbers called weights. Open-weights vendors like Meta publish that file, so it runs on your laptop or any host. Closed vendors keep it on their servers and sell answers through an API.
1 Guess first
2 Look
Two boxes: one locked, one a file
1 / 5your promptthe answerclosed · locked
GPT in ChatGPT, Claude, Gemini
the file stays on the vendor’s servers
you see
- your text
- the answer
- the price per token
you never see
- the file
- how many numbers it holds
- where it runs
model cards on Hugging Face, Ollama library, OSI definition 1.0 · checked 2026-10-22
A closed model is a locked box. Text goes in, text comes out; the model itself never leaves the vendor.
Too big for your memory? quantization · why small models are good: distillation · run one locally · where to use gpt-oss-20b free
A closed model is a box you can only talk to; an open-weights model is a file you can copy, so it runs on your laptop or any host, under the terms of its license.
The weights are the numbers training produced: billions of them, nothing else. Open weights means the vendor publishes that file. Open source, by the Open Source Initiative’s definition, also needs the training code and enough detail about the training data to rebuild it. Llama ships the file only, under Meta’s own license.
4 You are the model
Four jobs, four constraints. For each, pick where the model should run.
job 1 of 4 · pick where the model runs
Summarise private notes that must never leave your 16 GB laptop.
model cards on Hugging Face, Ollama library, OSI definition 1.0 · checked 2026-10-22
Pick where this job should run.5 Check
Why keep the file? OpenAI’s GPT-4 report cites “the competitive landscape and the safety implications” for not even sharing the model’s size. Selling answers by API is the business. Free tiers win either way: hosts like Groq and Cloudflare can serve open models for free because anyone may run the file.
7 Say it back
8 Mission · 5 min
On Groq’s free tier, ask openai/gpt-oss-20b: “In one sentence, what is a model’s weights file?” Then open huggingface.co/openai/gpt-oss-20b and find two things the API never showed you: the license and the parameter count. That same file runs on a 16 GB laptop with “ollama run gpt-oss:20b”.
Three quick ones
Drawn at random from this lesson's pool; different next time.
Done?
Go deeper
Sources
- https://huggingface.co/openai/gpt-oss-20b (model card, checked 2026-10-22): license apache-2.0; “21B parameters with 3.6B active parameters”; can “run within 16GB of memory”; “If you are trying to run gpt-oss on consumer hardware, you can use Ollama”
- https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct (checked 2026-10-22): “Llama 3.3 Community License Agreement”, “a custom commercial license”; over “700 million monthly active users” a licensee must request a license from Meta; gated (“You need to agree to share your Hugging Face information to access this model”); pretrained on “~15 trillion tokens of data from publicly available sources”
- https://huggingface.co/google/gemma-4-31b-it (checked 2026-10-22): License Apache 2.0, 31B parameters
- https://ollama.com/library/llama3.2, /gpt-oss, /llama3.3 (checked 2026-10-22): llama3.2:3b 2.0GB; gpt-oss:20b 14GB, Apache 2.0; llama3.3:70b 43GB (default 4-bit tag)
- https://opensource.org/ai/open-source-ai-definition (OSAID 1.0, checked 2026-10-22): requires data information, “the complete source code used to train and run the system” and the model parameters, under OSI-approved terms
- https://arxiv.org/abs/2303.08774 (GPT-4 Technical Report, §2 Scope and Limitations): “Given both the competitive landscape and the safety implications of large-scale models like GPT-4,” the report gives no details of architecture, size, hardware, compute, dataset or training method
- https://developers.cloudflare.com/workers-ai/platform/pricing/ (updated 2026-10-20, checked 2026-10-22): 10,000 Neurons per day free; @cf/meta/llama-3.3-70b-instruct-fp8-fast listed
- aifree.dev catalog: groq-free-tier verified 2026-10-12, models include openai/gpt-oss-20b and openai/gpt-oss-120b; cloudflare-workers-ai-free-tier lists @cf/meta/llama-3.3-70b-instruct-fp8-fast
- Weight values in the picture’s “inside the file” state are illustrative, labelled as such
Published Oct 22, 2026 · updated Oct 22, 2026