Open-weight models — where anyone can download the parameters and run, inspect, or fine-tune them on their own hardware — have quietly caught up with the proprietary frontier. I compared four of the leading ones to see how they actually differ.
The four contenders. Llama 4 (Meta) is the multimodal generalist, built on a Mixture-of-Experts design with a context window up to 10M tokens. DeepSeek V3/R1 proved frontier-level reasoning can be trained for around $5.6M and released openly. GLM (Zhipu) is agent-native — tuned for coding and tool use, and strongly bilingual. Phi-4 (Microsoft) is the small one: just 14B parameters, yet it beats models five times its size on math and STEM.
Open-weight isn't open-source. The word "open" gets stretched to cover both, but they are not the same thing. Open-weight means the trained parameters — the model's learned numbers, the actual result of training — are free to download, run, inspect, and fine-tune on your own hardware. What you are allowed to do with them, though, is still set by a model-specific license that can restrict commercial use, certain regions, or how you deploy the model. Open-source is a stricter bar: not just the weights, but the code, the architecture, and the training method, all released under permissive rights to use and modify them — the standard that licenses like MIT and Apache come closest to. That gap is why the license matters as much as the benchmark. DeepSeek, GLM, and Phi-4 ship under MIT and are effectively unrestricted; Llama 4 is freely downloadable but carries Meta's own community license, which adds conditions — leaving it open-weight, yet not truly open-source.
Two philosophies. Three of the four bet on scale plus Mixture-of-Experts — huge total capacity, small active compute per token. Phi-4 bets the other way: less data, but "textbook-quality" synthetic data. Both camps increasingly lean on synthetic data and post-training to earn their reasoning gains.
No single winner. Read the benchmarks with care — many are vendor-reported, and popular tests saturate quickly and leak into training. The honest conclusion is that the best model depends on the job: task type, budget, hardware, language, and how open the license must be. If you are weighing whether to run one of these yourself, that trade-off is exactly what my AI inference cost assistant is built to make visible.