Open weights
Run it yourself or pay forever — this is the real infrastructure decision
What are open weights
When a model is released with open weights, the trained model itself is publicly available. You can download it, run it on your own hardware, modify it, fine-tune it on your own data, and deploy it without calling anyone's API.
The analogy from CONTENT.md is precise: the recipe is public, but you still need the kitchen.
Open weights does not mean free to use in any context — licences vary, and commercial use restrictions are common. It means the trained parameters are not locked inside someone else's server.
The case for open weights
Control. Your model runs on your infrastructure. Your data never leaves your environment. No API dependency. No pricing changes that affect your product overnight. No usage policies that might conflict with your use case. No per-token bill that scales with your success.
For organisations with privacy requirements that prohibit data leaving their environment — healthcare, legal, government, financial services — open weights models are often the only viable option. For products where AI cost is the dominant operating expense, running your own model can be significantly cheaper at sufficient scale.
The case for knowing what it costs
The kitchen is not free.
Running a frontier-quality open weights model requires serious compute. A 70B parameter model like Llama 3.3 70B needs multiple high-end GPUs to run at reasonable speed. GPU infrastructure has capital cost and operational cost. You are responsible for uptime, performance, scaling, and security. You maintain it.
For most teams, a hosted API is cheaper than the engineering and infrastructure required to run open weights at production quality. The break-even point — where self-hosting becomes cheaper than API costs — is high. Most products never reach it.
How to think about this decision
Open weights is the right choice when: data cannot leave your environment, you need to fine-tune on proprietary data, or your token volume has crossed the break-even point.
Hosted API is the right choice when: your team does not have ML infrastructure expertise, your volume is below break-even, or iteration speed matters more than cost.
Open weights availability is tracked on every model page on sourc.dev. Verified March 2026.
Verified March 2026 · Source: Meta AI, Mistral AI, provider licence documentation