Three open-weight families
Llama (Meta), Mistral (Mistral AI), and Qwen (Alibaba) are the three most influential open-weight LLM families, each shipping a range of sizes from small models you can run on one GPU to large ones that rival closed frontier models. They are the default starting points when you want to self-host, fine-tune, or keep data inside your own environment. Because all three release new versions frequently, the right pick is less about a fixed ranking and more about matching a family’s strengths, sizes, and licence to your task — and then benchmarking the candidates yourself.
Performance and strengths
Llama is the most widely adopted and best-supported in tooling and the community, with strong general reasoning across its size range and an enormous ecosystem of fine-tunes and quantisations. Mistral punches above its weight on efficiency, popularised mixture-of-experts designs that deliver strong quality at lower active compute, and offers very competitive larger models. Qwen is a standout for multilingual work — especially Chinese and other Asian languages — while staying strong in English and code. On English benchmarks the large variants of all three are close; pick by ecosystem (Llama), efficiency (Mistral), or multilingual reach (Qwen), then verify on your own prompts.
Hardware and fine-tuning
Hardware needs scale with parameter count. Small models (roughly 7-8B parameters) run on a single GPU, particularly when quantised to 4-bit or 8-bit, which slashes memory at a modest quality cost. Mid-size models want a high-memory GPU or two; the largest need multiple high-end GPUs. All three families support efficient fine-tuning via LoRA and QLoRA, so you can adapt a model to your domain on relatively cheap hardware without retraining everything. Llama’s tooling maturity makes it the smoothest for first-time fine-tuners.
Licensing and which to choose
Licensing is decisive for commercial products. Mistral and Qwen ship many releases under permissive Apache 2.0, allowing broad commercial use. Llama uses Meta’s community licence — commercial use is permitted but with conditions, including a large monthly-active-user threshold and naming requirements. Choose Llama for the richest ecosystem and tooling, Mistral for efficiency and permissive licensing, and Qwen for multilingual and Chinese-language strength. In every case, confirm the exact licence for the exact model version, and benchmark your shortlisted sizes on your real workload before committing.