Skip to content

← All series & guides

Series

Open-source AI explained

10 parts

What open means for models: weights, licenses, and the homework of running things yourself. Not a free lunch.

Start with part 1 →

  1. 1 What open means: weights, license, not a free lunch "Open" in AI is three different claims: open weights you can download, open source under the OSI bar, and a free chat box. This part names the gap, why Llama-style licenses are not MIT, and why local is a control choice, not a free lunch.
  2. 2 Hosted open models vs download to your machine Hosted open models and a download to your machine are different places for the same weights. This guide splits Groq, Together, Fireworks, and chat UIs from Ollama, LM Studio, and llama.cpp, and it flags the trap where one app has a cloud toggle.
  3. 3 Model size, quantization, and will my laptop run it? Parameter counts are clothing sizes, quantization shrinks the file, and RAM decides whether the laptop can talk. A 70B Q4 on 16 GB is how you get a leaf-blower fan and a multi-minute first token.
  4. 4 Privacy reasons normal people care A local desktop runner can still ship your prompt to a host if cloud mode is on. This guide maps what leaves (text, files, metadata), when local is worth the heat, and when a $20 work chat is the honest privacy move.
  5. 5 Quality without benchmark theater A Slack rank is not a quality test. Run five real tasks from your week, score pass or fail, and ignore a 0.3 point LMSYS jump when the file you load is 4-bit and the board ran something else.
  6. 6 Safety: random models on the internet A GGUF from a random Hugging Face account is software you run on your machine. Prefer official orgs and known converters, read the model card, match file size and license, and skip mystery installers before you load anything.
  7. 7 When ChatGPT, Claude, Gemini, or Grok is simply easier Open weights are a download you can run, not a writing personality. Everyday drafts, tools, and memory still belong on a Plus or Pro chat around $20 as of writing. Local still wins for files that cannot leave, offline, huge volume, or learning the stack.
  8. 8 Open-weight vs open-source software vs free to use Open-weight, open-source software, and free to use are three different claims. A free chat box is a fourth. Write three ticket sentences, attach the license PDF, and let whoever signs contracts decide work use.
  9. 9 Reading a model card without getting lost Read a Hugging Face model card like a spec sheet before you download. Check publisher, license, chat versus base, and file size against your RAM. Skip the giant scoreboard on the first pass so you do not fill the disk with a 70B souvenir.
  10. 10 Chat APIs for open models: Groq, Together, Fireworks Hosted APIs for open weights bill by tokens and still send prompts off your machine. Separate Groq from Grok, set a spend cap, and treat retention pages as part of the job.