Closed vs open-weight vs self-hosted
The three ways to get a model: rent a closed one via API, use an open-weight one someone hosts for you, or run open weights on your own machines.
In plain terms
It's renting vs. leasing vs. owning a car. Closed API (GPT, Claude, Gemini): taxi - best quality, zero maintenance, pay per ride, no control over the engine. Open weights hosted by a provider (Llama, Qwen on Together/Groq): lease. Self-hosted: you own the car and the repair bills.
Why it matters
Teams differ on privacy (medical data can't leave the building), cost at scale, latency, and customization. No single option wins; 'which model, served how' is one of the first questions on any AI project.
How it works
Closed: you send data to the lab's servers; you get frontier quality and features first. Open-weight: labs (Meta, Mistral, Alibaba, DeepSeek) publish the weight files; anyone may run them - via a hosting provider's API (same convenience, cheaper) or your own GPUs. 'Open' means weights are downloadable, not that training data or recipe is public.
When you use it
Default: start closed (fastest path to quality). Consider open when: strict data residency, massive predictable volume where GPU rental beats per-token pricing, need to fine-tune deeply, or offline/edge requirements.
Common mistakes
- Self-hosting for 'privacy' by default - enterprise API tiers offer no-training guarantees; running GPUs well is a serious ops job.
- Benchmark worship: picking the leaderboard #1 instead of testing on YOUR task.
- Assuming open = free - GPU time for a big model often costs more than API calls until you have real volume.
Best practices
- Prototype on a frontier closed model first; optimize cost later once the product works.
- Keep a provider-agnostic wrapper around model calls so you can swap vendors in one place.
Try it yourself
Get a free API key from Groq (hosts open models at high speed), call Llama with the same prompt you used on a closed model, and compare quality, speed and price per million tokens.
Resources
- LMArena leaderboard Crowd-voted model rankings - useful signal, never the whole answer.