This catalogue of open-weight and open-source LLM families is a reading guide, not a trophy shelf. It lists notable downloadable families by openness class and licence, reviewed September 2026. It is not a best-of list and it does not publish llm.build scores. A row tells you how we class the family and whether a family page exists. A family page exists where a buyer would pick that base and needs the adaptation reading of the licence. Learn to read the table before you pick a name.
By Shen Pandi, Towards AGI. Reviewed September 2026. This page answers: open-weight and open-source LLM families.
How to read a row
Each row is a family you might meet as a downloadable base, not a ranking slot. Family is the name buyers already say. Organisation is who publishes the weights you will actually fetch. Openness is a class: open-source when published artefacts are intended for reproduction, open-weights when you can download parameters subject to the card, restricted-weights when the grant is narrower than a typical commercial adaptation. Licence is a short pointer, not the agreement. Size band is a serving hint, not a quality hint.
A clickable family name means we own an adaptation reading at that URL. An unlinked name means the family belongs in the landscape so a brief can recognise it, but we have not yet taken a licence-and-adaptation query we can stand behind. Do not treat an unlinked row as a snub. Do not treat a linked row as a winner. The table is how we talk about the landscape without a thin page per release.
What a row is for, and what it is not
Use a row to answer four reading questions. Is this a name we expect in a 2026 brief? Which openness class did we assign from public sources? Is there a family page that owns the surviving duties after adaptation? Is the size band in the range you actually operate? Those are catalogue questions. They are not “which model should we buy?”
Do not use a row as a licence exhibit in a board pack. The short licence string is a pointer. The family page, or the checkpoint file itself, is the exhibit. Do not use a row as a serving plan. “8B-class” is not a capacity study of your sequence length. Do not use a row as a quality claim. We did not measure these families on your book.
Family: a name a buyer already says, not a release we ranked.
Openness: a class, not a slogan you can paste into an RFP.
Licence: a pointer. The family page or the checkpoint file is the exhibit.
Href: we own an adaptation reading. No href: landscape only, for now.
When a family gets a URL
A family appears in the table if a commercial team is likely to meet it as a downloadable base. A family gets its own URL when there is a licence or adaptation query we can actually own: Community Licence after a LoRA, Apache on a current Qwen3 card versus earlier generations, Gemma Terms versus a later labelled card, Mistral 3 Apache versus a revenue-clause older card, DeepSeek R1 MIT versus V3 Model License, OLMo as a reproduced stack. New checkpoints do not get their own URL until that query exists.
That rule is why Pythia can sit in the table without a spoke, and why Granite, GLM, Kimi and MiniMax can appear as names you will hear without a page that pretends we have taken their grant. A URL per checkpoint goes stale. The table is how we stay complete without farming thin pages. If Search Console or a sales brief later shows a query we can own, the family earns a spoke. Until then, read the row as landscape.
The catalogue note is the honesty line
Listed because a buyer might meet the family as a downloadable base, not because llm.build ranked it. Licences and openness class are read from the named public sources on the family page or the checkpoint card. Confirm the card before you train. Reviewed September 2026.
That note is doing three jobs. First, listing is not ranking. Second, the short licence string was read from named public sources on the family page or the checkpoint file — not invented here. Third, the review month is part of the artefact. A table without a date becomes folklore. If a card moves after September 2026, the family page moves. This index moves when the landscape rule moves, not when a listicle adds a name.
Open weights are not open source
Most rows are open weights: you can download parameters and, subject to the card, fine-tune them. Open-source AI, as the OSI defines it, also needs the preferred form to modify the system. Read the openness comparison if that distinction is the procurement question. This index will not reprint OSAID. It will tell you which rows are trying to be that exhibit (OLMo, Pythia) and which rows are downloadable weights with a software licence on the file.
A slide that says “we picked an open-source model” after someone pointed at this table is a vocabulary error. Point them at the openness comparison, then at the family page if one exists. Do not let this index become the place that error gets laundered.
Small models live in this table
Phi, small Gemma, small Qwen, small Llama, SmolLM, TinyLlama, MiniCPM and OpenELM are SLM-class entries here. The commercial `/slm` page is still the place to decide whether a smaller model is the engagement. We did not ship `/models/slm` so those queries would not compete. Compact size is a serving conversation. It is not a second ranking URL.
If the brief is “we need it small”, start at `/slm` for whether shrinking is the job, then come back here to see which families even have a compact band. Do not pick TinyLlama from this table because the name sounds cheap. Do not pick a 70B-class row because a listicle said it won a public exam.
Practice: counsel used open source, the table has three kinds of row
A wording lead reads the catalogue before anyone books a base
insurance
Counsel wrote that the extract project must use an open-source model. A wording lead opens this index instead of a blog list. They read openness class, not family fame. OLMo is linked because we own the reproduced-stack reading. Llama is linked because we own the Community Licence after adaptation. Pythia sits unlinked: landscape, not a spoke we have taken. The lead does not pick a winner. They write two sentences: we need OSAID-shaped artefacts, or we need downloadable weights counsel will accept under a community grant.
If the first sentence is true, they open the OLMo page, not this table’s licence cell. If the second sentence is true, they open Llama or the openness comparison and stop calling the project open source. Unused test documents for the extract still have to exist either way. The catalogue did not score the wording family. It stopped a misnamed base.
Practice: a catalogue owner needs a compact band, not a trophy
Merchandising reads size band and href, then leaves the table
retail
A retailer wants attributes from supplier text on hardware they already operate in the 8B-and-under band. They open this index to see which families even publish that band and which of those have a page. Phi is linked because we own the SLM-as-base reading. Several compact names are unlinked. They do not treat the unlinked names as worse. They treat them as “no adaptation essay yet”.
They then leave the table. The next URL is `/slm` if the question is whether to shrink at all, or a family page if a named base is already in the brief. They do not copy every licence cell into a steering pack. They do not ask us which compact row “wins attributes”. Unused SKUs still have to exist before anyone trains.
How we choose what is listed
We include true open-source stacks (OLMo, Pythia), widely used open weights (Llama, Qwen, Mistral, Gemma, Phi, DeepSeek), compact SLMs, and other names that show up in 2026 briefs (Granite, GLM, Kimi, MiniMax). We include restricted-weight rows such as OpenELM, Command and BLOOM so a buyer who meets those names does not think we forgot them. Forgetting is worse than listing a narrow grant.
We do not list a family because we like it. We do not unlist a family because a licence is awkward. Awkward grants are why some rows exist. If your brief named a family that is not here, tell us. Absence is either “we have not seen that name in briefs yet” or a gap we should close in the next review month.
Do not reprint the licence from the cell
This index will not walk Llama’s attribution rules, Qwen3’s Apache file, Gemma Terms, Mistral’s split, or DeepSeek’s R1-versus-V3 grant. Those readings have URLs. Pasting them here would make three pages rot together. The licence cell is a reminder to go read. If you need the surviving duties after a LoRA, open the family page.
If there is no family page, the reminder is sharper: you will read the checkpoint file yourself, or you will ask us to take that query. Do not invent a duty list from the short string in the table. Short strings collapse distinctions the cards do not collapse.
Decision prompts, not scores.
When
Then
You need the adaptation reading of a named base
Open the family page if the name is linked. Do not stop at the licence cell.
Counsel said open source and you are staring at this table
Read openness class, then the openness comparison. Do not paste a row into the RFP.
The family is listed but unlinked
Treat it as landscape. Read the checkpoint file or ask us to take the query.
You wanted a winner among families
Wrong index. Write the task, the inference boundary and unused test documents first.
Href is a query we took, not a medal
Llama, Qwen, Mistral, Gemma, Phi, DeepSeek and OLMo have pages because a buyer already arrives with a licence-and-adaptation question we can write down. Pythia sits next to OLMo in the table because it is the other named reproduction stack, not because we owe it a second essay this month. Granite, Falcon, Yi, GLM, Kimi and MiniMax appear because briefs already say those names. Appearance is recognition. A URL is a query we accepted.
If your steering pack only lists linked families, you have turned this index into a shortlist we did not write. Read the unlinked rows. Ask whether the named base is actually in the brief. If it is, and there is no spoke, the next step is the checkpoint file plus a question to us — not a silent downgrade of that family.
Restricted-weight rows are still part of literacy
OpenELM, Command and BLOOM sit here so a buyer who meets those names does not think the catalogue forgot them. Restricted-weight means the grant is narrower than a typical commercial adaptation. That is useful information. It is not a moral grade. A team that needs a research probe can still evaluate a restricted row. A team that needs to ship a closed product should read that grant before they treat the row like Apache.
Do not skip these rows because they complicate the story. Complication is the job of a catalogue. A table that only listed Apache-shaped names would be a marketing page. This one lists what you will actually hear, including awkward grants.
What this index does not replace
It does not replace a family page. It does not replace the openness comparison. It does not replace `/slm`. It does not replace a data-readiness review. It does not replace the commercial fine-tuning page. Buyers who bookmark only this URL still have to pick a next URL. That is the point of an index.
If you arrived from a search for a single family, skip the sermon and use the href. If you arrived from a search for “open-weight and open-source LLM families”, stay here long enough to learn the columns, then leave with one next URL. Lingering to compare every row is how this page gets misused as a leaderboard.
Catalogue
Notable open-weight and open-source families as of September 2026. Not a ranking.
We do not publish a winner. We recommend a base after the task, licence, inference boundary and unused-test plan are written down.
Why is there no page for every family?
A URL per checkpoint goes stale. The table is how we talk about the landscape. Family pages are for licence-and-adaptation queries we can actually own.
Do you measure these models?
Not on this page. Any number you see elsewhere must be cited to a named public source with a date. Customer results appear only after a scoped evaluation.
What does a linked family name mean?
We own an adaptation reading at that URL. It is not a quality badge. An unlinked name is still a family you may meet; we have not taken that query yet.
Can I paste the licence cell into a board pack?
No. The cell is a pointer. Use the family page or the checkpoint file as the exhibit. Short strings collapse distinctions the cards do not collapse.
Where do small models live?
In this table, as SLM-class rows. The commercial `/slm` page is where you decide whether shrinking is the engagement. We did not ship a second ranking URL.
Is this the page that explains open weights versus open source?
No. This index tells you how to read openness class in the table. The comparison page owns the definition.
What if our brief named a family that is not listed?
Tell us. Absence is either a name we have not seen in briefs or a gap for the next review month. Do not treat silence as a ranking.
Pick a family only after you pick a task.
Use the table to see openness and licence. Then bring one workflow and where inference must run. We will say whether a family page, a readiness review or a commercial engagement is next.