Module 4 of 5 · Intermediate · 7 min read
Open weights vs closed models
When to self-host, when to use an API, and what 'open' really means.
- 1Open weights let you run, inspect and fine-tune a model yourself.
- 2Closed frontier models still lead on the hardest tasks.
- 3Licences vary: read them before building a product.
01What 'open' means
Open-weights models publish their trained parameters so anyone can download and run them. Few publish their training data or full recipe, so "open source" in the strict sense is rare.
Licences range from permissive (MIT, Apache 2.0) to restrictive (usage limits, non-commercial clauses). DeepSeek V4.1 Flash is MIT-licensed; others have custom terms.
02Why teams choose open
Control and cost are the big reasons.
- Data never leaves your infrastructure.
- Fine-tune deeply for a niche domain.
- No vendor lock-in or surprise deprecations.
- Very low marginal cost at high volume, if you have the GPUs.
03Why teams choose closed
Closed frontier models still lead on hard reasoning, long-horizon agents and computer use. APIs handle scaling, safety filtering and upgrades for you, and enterprise agreements cover data handling.
Many teams use both: a closed flagship for hard tasks and an open model for high-volume or sensitive ones.
Finished reading?
Mark it complete to track your progress, then test yourself with the Models & Labs quiz.