Open-weight models
Open-weight models are AI models whose trained weights, the billions of numbers that make up the model, are published for anyone to download and run. You can run one on your own servers or with any hosting provider you like, instead of only through the maker’s API. Llama, DeepSeek, Qwen, Mistral and OpenAI’s gpt-oss models are examples.
Open weights are not the same as open source
Releasing the weights doesn’t mean releasing everything. Most open-weight models come without their training data or the full code used to train them, so nobody can rebuild them from scratch. Licenses vary too. Some allow almost any use, while others restrict large commercial users or rule out certain uses. Read the license before you build on one.
Why it matters
- Cost: many hosts compete to serve the same open model, which pushes prices down. On a lot of everyday work, open models now come close to the leading closed models for a fraction of the price.
- Control: you choose who hosts the model and where your data goes, and you can keep using the same version for as long as you want.
- Less lock-in: a closed provider can change prices, rate limits or which models it offers. With open weights, someone else can always serve the model.
- Customization: you can fine-tune the weights on your own data.
The trade-offs
The best closed models still lead on the hardest tasks, and some work is worth paying more for. Running a model yourself also means owning the serving, scaling and security work, which is why many companies use a hosted provider for open models rather than buying their own GPUs.
How elvex relates
elvex Enterprise runs on open-weight models and sends each task to a lower-cost model where it does the job, keeping premium models for the work that needs them.