Open-Source LLMs
Whether you can download the weights decides who controls the model: you can fine-tune it, run it on your own GPUs and keep it offline — in exchange for running inference, upgrades and incident response yourself.
The open-versus-closed decision is rarely about which model is stronger. It is a combination of four premises: whether data may leave the network, whether the team can operate infrastructure, whether call volume will stay high, and how far the current option lags.
Where data is sensitive or regulated, open weights plus private deployment is close to the only answer. Where volumes are low and requirements shift weekly, the flexibility of a hosted API is worth more. The hidden cost sits in operations: inference serving, performance tuning, version upgrades and incident response — open source does not end at the download.
How much is actually opened also varies. Qwen3-Omni ships Instruct, Thinking and Captioner variants under an Apache-2.0 license; MiniMax M2.1 uses a Modified-MIT license, targets coding and agent workloads, and publishes SWE-bench Verified and similar numbers; other releases ship weights only, without training data or with usage restrictions. Reading the license before signing anything matters more than reading the leaderboard.
