A map of open-source models

The video is in Italian; this page is the written version in English.

In the open-weights world there are only a few families that really matter, each with a recognisable philosophy. Knowing who does what saves you from trying twenty models at random.

Qwen (Alibaba)

The best quality-to-size ratio, and strong multilingual behaviour. Sizes cover everything from 0.6B to a 235B MoE, and the Apache 2.0 licence is the most permissive around.

Qwen 3’s distinguishing feature is switchable reasoning: you can ask the model to think step by step or to answer immediately, with the same weights. There are also code-specialised variants.

Llama (Meta)

The industry reference model — not necessarily the best on every metric, but the one the largest ecosystem of tools, adapters and tutorials grew around. If a project supports only one model, it is almost always Llama.

Sizes run from 1B to 405B, including variants that can see images. The licence is permissive but not Apache: there is a clause about very large companies, worth reading in a commercial setting.

DeepSeek

It brought explicit reasoning into the open-weights world. The R1 models show their thinking before answering and hold their own against closed models on hard problems. They use MoE architectures, so they carry many total parameters but activate a fraction: enormous on paper, surprisingly practical to run.

Mistral (France)

European models, efficient, with strong results in small and mid sizes and generally clean licences. A sensible choice when data provenance and jurisdiction matter.

Gemma (Google)

Derived from the research behind Gemini, these aim at modest sizes with good quality per parameter. Good candidates for running something on limited hardware.

How to read the licences

The part everyone skips and later regrets.

Licence	In practice
Apache 2.0 / MIT	Do what you like, commercial included
Llama License	Permissive, with conditions for very large platforms
“Research only” licences	No commercial use: read before building on it

Watch derived models too: a fine-tune inherits the constraints of its base, however much it was retrained.

Where to find them

The gathering point is Hugging Face, where every model has a card with sizes, formats and licence. For local use what matters most is availability in GGUF format, the one the tools in the Local Setup section can read.

How to read a model name

The names look like arbitrary strings and in fact say almost everything. Reading them saves useless downloads.

Qwen3-8B-Instruct-GGUF-Q4_K_M
 │    │   │        │    │
 │    │   │        │    └─ quantization: 4 bits, K_M variant
 │    │   │        └────── file format: what runs it
 │    │   └─────────────── variant: trained to follow instructions
 │    └─────────────────── size: 8 billion parameters
 └──────────────────────── family and generation

Part	What it tells you
No suffix, or Base	Raw model: does not converse
Instruct, Chat	Ready to use
Coder, Math, Vision	Specialisation
GGUF, MLX	Format, i.e. which program runs it
Q4_K_M, Q8_0	How compressed it is
Distill	Learned from a larger model

Two checks before downloading: the file size, which matters for memory far more than the parameter count; and who uploaded the repository, because third-party conversions are often excellent but licence and quality remain whatever the uploader states.

Derivatives: distilled, merged, retrained

Many names are not original models but derivatives, and expectations differ.

Type	How it is made	What to expect
Distilled	Trained on a larger model’s outputs	Inherits style, not fully the capability
Merged	Average of two models sharing a base	A compromise, unpredictable
Community fine-tune	Retrained on specific data	Great in its domain, often worse outside

A name like R1-Distill-Qwen-8B reads as: the base is Qwen 8B, the extra training comes from a larger model’s reasoning. It is not “that model in miniature”, it is a Qwen that learned its way of reasoning.

Prudent rule: community derivatives should be tried on your own case. The enthusiastic scores that accompany them are often obtained on tasks very close to their training data.

In short

Family	Why pick it
Qwen	Quality per size, strong multilingual, Apache licence
Llama	Largest tooling ecosystem, industry reference
DeepSeek	Explicit reasoning, efficient MoE
Mistral	European efficiency, good mid sizes
Gemma	Small models with good output
Model name	Tells variant, format and compression: read it before downloading
Derivatives	Distilled and merged models need testing, not trust