Open-source AI models · Global AI Product Index

How to choose

Who this is for. Builders and small teams choosing an open-weight family to self-host, fine-tune or evaluate, usually because they need to run the weights themselves or check a licence before committing to one.

Who this is not for. People who only want a hosted chat assistant and will never touch the weights, teams that need a vendor support contract with one accountable provider, and anyone comparing models purely on benchmark scores.

Read the licence recorded for the family before shipping on it: some entries are permissive, others carry conditions or non-commercial limits.
Match size to hardware: the flagship group needs serious GPU capacity or a hosted endpoint, while the small group targets local running.
Decide whether weights alone are enough or you also need data and training code: open weight is not the same as open data, and each record notes what is actually published.
Check what enterprise-oriented families add on top — retrieval, tool use or support arrangements — since the same weights can hide a very different deployment story.

Flagship open-weight LLMs

Large general-purpose families that other models and products build on, spanning chat, reasoning, coding and multilingual work. They differ from the small group after them, which trades scale for low-cost local running, and from the enterprise group that follows, which layers retrieval or support on top.

AI products
Name	Company	Category	Business model	Launched	Updated
Llama Running models yourself	Meta	Open Models	Open (open weights/source)	2023-02	2026-09-04
Mistral Running models yourself	Mistral AI	Open Models	Open (open weights/source)	2023-09	2026-10-13
Qwen Running models yourself	Alibaba	Open Models	Open (open weights/source)	2023-08	2026-10-13
DeepSeek Running models yourself	DeepSeek	Open Models	Open (open weights/source)	2023-11	2026-10-13
GLM Running models yourself	Z.ai (Zhipu AI)	Open Models	Open (open weights/source)	2023	2026-09-10
Baichuan Running models yourself	Baichuan Intelligence	Open Models	Open (open weights/source)	2023-06	2026-09-10
Yi Running models yourself	01.AI	Open Models	Open (open weights/source)	2023-11	2026-09-10

Small and efficient models

Compact families positioned to run locally or cheaply rather than at maximum scale. Unlike the flagship group before them they are not general-purpose bases, and unlike the enterprise group after them they come without a vendor deployment or support story attached.

AI products
Name	Company	Category	Business model	Launched	Updated
Gemma Running models yourself	Google DeepMind	Open Models	Open (open weights/source)	2024-02	2026-09-10
Phi Running models yourself	Microsoft	Open Models	Open (open weights/source)	2023	2026-10-09
OLMo Running models yourself	Allen Institute for AI	Open Models	Open (open weights/source)	2024-02	2026-10-10

Enterprise-focused models

Families positioned for enterprise assistants, retrieval or licensed support rather than for general chat. Unlike the small group before them, size is not the defining trait, and unlike the flagship group they are usually presented alongside a vendor's cloud deployment, a support route or a specific licence.

AI products
Name	Company	Category	Business model	Launched	Updated
Command R Running models yourself Questions over your own documents	Cohere	Open Models	Open (open weights/source)	2024-03	2026-10-09
Granite Running models yourself	IBM	Open Models	Open (open weights/source)	2023-09	2026-09-10
Nemotron Running models yourself	NVIDIA	Open Models	Open (open weights/source)	2024	2026-10-10
Jamba Running models yourself	AI21 Labs	Open Models	Open (open weights/source)	2024-03	2026-09-10
DBRX Running models yourself	Databricks	Open Models	Open (open weights/source)	2024-03	2026-09-10
Falcon Running models yourself	Technology Innovation Institute	Open Models	Open (open weights/source)	2023-05	2026-09-10

Your next step

Every link above keeps the evidence attached: the index shows each value with its source and the day it was checked.

Frequently asked questions

Does this guide rank or benchmark these models?

No. It groups indexed families by scale and positioning and links to sourced records. It publishes no scores, no benchmark results and no leaderboard ordering, and nothing here is a recommendation.

Is open weight the same as open source?

Not necessarily. Open weight means the weights are downloadable; training data, training code and licence terms vary between families, and the index records which of those a release actually publishes.

What are the limits of this guide?

Coverage stops at families already indexed here with a sourced record, so many open-weight releases are absent, and both model generations and licence terms change faster than a static guide can follow.

Where do the facts come from?

From each product record in this index, which cites official and independent sources and a last-reviewed date. Nothing in this guide goes beyond what those records state.

Can I use these weights commercially?

It depends on the family. Licences recorded in the index range from permissive terms to vendor community licences with conditions, and at least one is limited to non-commercial use, so read the individual record before planning on it.

Related comparisons

Qwen2.5 vs DeepSeek-V3 vs Mistral Small 3.1

Sources and record dates

Record dates describe index updates, not independent product tests. Check linked documentation for current features and pricing.