Open weights means a model’s trained parameters — the billions of numbers that are the model — are published for anyone to download, run on their own hardware, and usually fine-tune. The opposite is a closed model like Claude Opus 4.8 or GPT-5.5, reachable only through the vendor’s API, where the weights never leave the lab.
Open weights ≠ open source
The terms get conflated constantly, and the difference is contractual, not academic. Open weights tells you the files are downloadable. Open source describes the license — and model licenses range from genuinely permissive to restrictive:
- True open source (MIT, Apache-2.0): use, modify, commercialize freely. GLM-5.2 ships under pure MIT; ACE-Step under Apache-2.0.
- Modified / conditional licenses: mostly permissive with carve-outs — Kimi K2.7 Code uses a Modified MIT.
- Community licenses with commercial restrictions: downloadable, but usage caps or revenue thresholds apply — MiniMax M3’s license restricts commercial use.
Training data and code usually stay private even for “open” models, which is why the OSI’s stricter Open Source AI Definition exists. Read the license before building a business on the weights.
The actual trade-off
Open weights buy you control: data never leaves your infrastructure, no vendor can deprecate or throttle you, per-token cost at scale can undercut APIs by 5–10×, and fine-tuning is unrestricted. Closed models still tend to hold the capability frontier — though the gap narrowed dramatically by 2026, with GLM-5.2 beating GPT-5.5 on SWE-bench Pro. What the sticker price hides: self-hosting a 700B-parameter model means serious GPU capacity, ops expertise and per-hour costs that only pay off at sustained volume. Filter any model directory listing by license before assuming “open” means “free.”