Open Model Families: Who Ships What
5 min read · updated August 3, 2026
Any page listing the current best open models is wrong by the time you read it. What does not go stale is the structure: which organisations publish weights, why they do it, what sizes they publish in, and what sort of licence each tends to attach.
Why a list of names goes stale
Release cadence in this ecosystem is measured in weeks. A named recommendation ages badly in both directions: the model you named is superseded, and the one that replaced it may sit under a different licence with different obligations. Worse, the same family name across two releases can mean two materially different licences, so even “that family is permissive” needs re-checking at each version.
So this page describes the terrain and hands you a procedure. The terrain moves slowly; the coordinates move fast.
Who publishes, and what they optimise for
Motivation predicts behaviour, including how likely a publisher is to keep publishing and how permissive they are likely to be.
- Large platform labs — Meta’s Llama family, Google’s Gemma. Open weights are a strategic play: commoditise the layer, build the ecosystem, keep the frontier model closed. Typically a bespoke community licence, broad commercial permission with conditions attached.
- Model-first companies — Mistral AI, DeepSeek, Alibaba’s Qwen, and several others. Open weights build reputation and adoption while a paid API and enterprise licensing carry revenue. Licensing varies release by release, often deliberately: permissive on the smaller models, more restrictive on the largest.
- Enterprise vendors — IBM’s Granite, Nvidia’s Nemotron, Microsoft’s Phi. Aimed at customers who need clean licensing and documented provenance more than they need the top of a leaderboard. Historically the most reliably permissive corner of the ecosystem.
- Open-research organisations — AI2’s OLMo, EleutherAI, and other fully documented releases. Data, code and intermediate checkpoints as well as weights. Rarely the strongest models, uniquely valuable if you need to know what went in.
- The fine-tuning community. Not a lab, but a very large share of what is downloadable: instruction tunes, merges, domain adaptations and quantised repackagings built on the bases above. Inherits both the capability and the licence of its base.
The size classes
Sizes cluster, because they are chosen to fit hardware. These bands have been stable for a while and are the right unit for planning:
| Class | Description |
|---|---|
| 0.5B - 3B | On-device and edge. Classification, extraction, routing, autocomplete, draft models for speculative decoding. Runs on a phone or a CPU. Not a general assistant. |
| 7B - 9B | The workhorse band and the most competitive one. Fits at four-bit in single-digit gigabytes, genuinely useful across summarising, rewriting, structured extraction and everyday code. Where most local deployments end up. |
| 12B - 32B | A real step in reasoning and instruction adherence. Wants a serious card at four-bit. The best quality most people can run on one machine. |
| 70B - 120B dense | Multi-GPU, or heavy quantisation on a large unified-memory machine. Closes much of the gap on hard tasks; a substantial infrastructure commitment. |
| Large sparse (MoE) | Very large total parameter counts with a small active fraction. Compute economics of a mid-size model, memory requirements of a very large one — which is why these are usually rented rather than owned. |
Licence character by publisher
Expectations, not assertions — verify against the release you are downloading, and verify again at the next version.
- Apache-2.0 or MIT appears across much of the ecosystem and is common from enterprise vendors and research organisations, as well as on many individual releases from model-first companies. This is the case with no strings.
- Bespoke community licences are the norm from the large platform labs. Expect an acceptable-use policy incorporated by reference, redistribution conditions, naming and attribution rules for derivatives, and in some cases a very high user-count threshold above which separate permission is required.
- Split licensing within one family is common: smaller models permissive, largest model under a research or commercial-negotiation licence. Never infer the licence of one size from another.
- Research-only releases are usually clearly marked and often pair with an available commercial licence. Fine for evaluation; a hard stop for production until you have bought one.
None of this is legal advice. Read the LICENSE file and any policy it references in the exact repository you download from.
Refreshing this yourself
Fifteen minutes, and you have a current picture rather than a remembered one:
- Take a current leaderboard and filter to open weights, then filter again to the size class that fits your memory budget. That is your candidate set — three or four models, not thirty.
- For each, open the model card and read the licence file, not the summary line. Note the licence name and the date you read it.
- Check what the publisher shipped besides weights: base and instruction-tuned variants, official quantisations, a documented chat template.
- Check whether anyone serves it, which is a decent proxy for whether the ecosystem has settled on it and whether tooling support is real.
- Run your own evaluation set against the survivors. The leaderboard produced the shortlist; it does not produce the decision.