Any ranking of these families has a shelf life of about a quarter. What does not change that fast is the *shape* of each family: what sizes exist, what licence they ship under, what tooling assumes them, and what they were optimised for. Choose on the durable properties, then measure.
Qwen (Alibaba) — the widest ladder
The most complete range of sizes anywhere: roughly 0.6B up to 235B, dense models and mixture-of-experts, with dedicated coder and math variants. The main Qwen3 line is Apache 2.0. Multilingual coverage is broad and Chinese-English is particularly strong.
Why it matters: there is a Qwen for the machine you actually have. You can prototype at 32B on a rented GPU and ship the 4B on a laptop with the same prompt format and the same tokenizer family. If you do not know where to start, start here.
Llama (Meta) — the ecosystem
Rarely the top of any chart now, and still the model most tools assume. Quantizations exist for everything. Fine-tuning recipes, serving configs, LoRA adapters, papers, Stack Overflow answers, and half the inference stack were written against Llama first.
Custom community licence with the naming, attribution and user-cap clauses from the previous lesson. Pick Llama when integration surface and the sheer weight of prior art matter more than a benchmark position.
Mistral (France) — permissive and practical
Most of the line is Apache 2.0. Efficient dense models, a strong 12B and 24B class, good European language coverage, and reliable function calling. Historically the family that made small models feel usable.
One caution: not every Mistral release is Apache. Codestral shipped under a non-production licence. Check per model rather than assuming the family posture.
DeepSeek — research in public, at frontier scale
The V3 mixture-of-experts line and the R1 reasoning line, with unusually detailed papers describing how they were trained. R1 is MIT, which is remarkable for a model at that capability level.
The honest caveat: the flagship is around 671 billion parameters with roughly 37 billion active. You are renting it, not running it. What most people actually run are the R1 distills into Qwen and Llama bases — which are good, are weaker than the teacher, and carry the base model's licence.
Gemma (Google) — quality per byte, with strings
Gemma 3 covers 1B, 4B, 12B and 27B, with vision on the 4B and up and long context on the larger sizes. Pound for pound this family tends to be excellent at the small end, which is exactly where most people are constrained.
The strings are the licence: custom terms, a prohibited use policy that Google can update, and an obligation to pass the restrictions downstream.
Phi (Microsoft) — synthetic data, small and sharp
Built on heavily curated and heavily generated training data rather than raw web scale. Phi-4 is 14B under MIT, with smaller and reasoning-tuned siblings. It scores strikingly well on reasoning and maths benchmarks for its size.
The known trade-off is worth stating plainly: a model trained mostly on textbook-shaped data has less long-tail world knowledge and can be more sensitive to prompts that do not look like its training. It rewards a narrow, well-specified task and punishes open-ended chat more than its benchmark numbers suggest.
Worth knowing about
- OLMo (AI2) — data, code, checkpoints and weights all released. The family to use if you need to study or defend how a model was trained.
- Granite (IBM) — Apache 2.0, enterprise framing, decent tool use, sizes tuned for cheap serving.
- SmolLM (Hugging Face) — very small, fully open, honest about being small.
- Kimi, GLM, MiniMax — large mixture-of-experts models with strong agentic and coding claims; typically too big to self-host.
- Nemotron (Nvidia) — re-tuned derivatives of other bases, optimised for Nvidia serving.
How to use this map
None of these holds a durable lead. What is durable: licence posture, the size ladder, ecosystem depth, language coverage, and house style. Narrow on those. Then test the two or three survivors on your own examples, because the family reputation tells you nothing about your task.
Before you move on