What are the best open-source LLMs?

As of July 2026, the leading open-weight LLMs are DeepSeek V4 (MIT-licensed), Alibaba's Qwen family, Z.ai's GLM, Moonshot's Kimi, and Meta's Llama — with the Chinese labs currently setting the pace at the top of the open leaderboards.

One correction your smart friend owes you: almost none of these are truly open source. You get the weights — the trained model file — but usually not the training data or the full training pipeline. "Open-weight" is the accurate term, and for most practical purposes it's what matters: you can download the model, run it on your own hardware, and fine-tune it.

How to actually pick one:

• Check the license first. MIT (DeepSeek V4) and Apache 2.0 (many Qwen releases) are safe for commercial use; some models ship custom licenses with usage restrictions buried in them.
• Match size to hardware. The frontier open models are enormous — great on rented GPUs, hopeless on a laptop. Smaller releases like Google's Gemma or mid-size Qwen models are the local-machine picks.
• Check task-specific leaderboards. The best overall open model changes roughly monthly, and the best one for coding agents is often a different model entirely.

The gap between open-weight and closed frontier models has narrowed to months, not years. For current standings, check the live model tracker, and see what open-source LLM really means for the licensing fine print.