Open weights vs open source
A trained model is, at bottom, a very large file of numbers — its weights or parameters. A vendor that publishes that file has released an open-weight model: you can download it, run it, and usually fine-tune it, within the terms of its licence.
Open-source AI is a stricter claim. The Open Source Initiative's Open Source AI Definition (version 1.0) requires three things: the model parameters, the complete source code used to train and run the system, and enough information about the training data that a skilled person could build a substantially equivalent system. Most models people call open source publish the first and not all of the rest, which makes them open-weight.
The licence decides what open means in practice. On Hugging Face, Alibaba's Qwen3.5-397B-A17B is published under Apache 2.0, and Zhipu's GLM-5 and DeepSeek's V3.2-Exp under MIT — both permissive. MiniMax's M2.1 is published under its own licence, which is the case where the terms need reading before commercial use.
Three ways to run one
- From the vendor's API. Many open-weight vendors also sell the same model per token — Qwen, GLM, DeepSeek and MiniMax all have pages in our index with their list rates.
- From a third-party host. Inference providers serve popular open models at their own per-token prices, which can undercut or exceed the vendor's own.
- On your own hardware. You pay for GPUs, not tokens. The largest open models need multiple data-centre GPUs; smaller ones run on a single card or a laptop.
When self-hosting is cheaper
Per-token APIs charge only for what you use; owned or rented GPUs cost the same whether busy or idle. Self-hosting therefore pays when traffic is high and steady enough to keep the hardware busy, or when data rules require the model to run inside your own environment. For spiky or modest traffic, an API — the vendor's or a host's — is usually cheaper, because the provider spreads idle capacity across many customers.
Open weights also change the risk picture. An API model can be retired on the vendor's schedule; a copy of open weights keeps working for as long as you run it. Where long-term reproducibility matters — a regulated process, a research baseline — that can be worth more than the per-token price.
Things to check first
Read the licence, not the label: permissive licences such as MIT and Apache 2.0 allow commercial use broadly, while custom licences can restrict use cases or scale. Check what was released — some vendors publish weights for a smaller sibling while keeping their flagship API-only. And remember that the tokenizer ships with the weights, so token counts on an open model can be reproduced exactly offline, unlike on a closed one.