Llama
Meta's Llama family is the best-known option you can download and run yourself, which matters for privacy, offline use and cost control. Each version ships under its own licence.
Summary is our draft- Sources checked
- 11 Nov 2026
- Our testing
- Not started
Current lineup
Llama 4
Natively multimodal models with a mixture-of-experts architecture
Downloadable weights
Llama 3 (3.1, 3.2, 3.3)
No provider description captured
Downloadable weights
Llama 2
No provider description captured
Downloadable weights
Verified 12 Nov 2026Meta: Llama models and licences (opens in a new tab)
Which one for you
Each line says whether it comes from the provider or is our own judgment from the checked facts. None of it is from our own testing yet.
Data that must stay on your own machine
Run a Llama model locally with a tool such as Ollama. Nothing you type leaves the computer.
Our judgment
Learning how AI works under the hood
Running a small open model yourself teaches more about context, speed and hardware limits than any hosted app.
Our judgment
Business use
Read the licence and use policy for the exact version first. Terms differ between versions.
Our judgment
Key facts
- Licences
- Llama 4 License, Llama 3.3 License, Llama 3.2 License and earlier, each with a use policy
- Source, 12 Nov 2026 (opens in a new tab)
- Context window
- Per-model detail: not yet checked
- Hardware needed
- Depends on model size; to be documented: not yet checked
By task
We haven't run our own tests yet, so this shows only what the provider claims and the test every family will get. Results will appear here, labelled as ours.
Coding
The provider saysNo specific claim capturedThe test we’ll runRefactor a 400-line TypeScript module with an existing test suite, then fix one failing test.
Writing
The provider saysNo specific claim capturedThe test we’ll runTurn a rough 600-word draft into a 300-word brief without losing any of its facts.
Research
The provider saysNo specific claim capturedThe test we’ll runSummarise three supplied PDFs and cite the page for every claim.
Reasoning
The provider saysNo specific claim capturedThe test we’ll runPlan a five-step project with dependencies and explain which step blocks which.
Pricing and access
Pricing
- Weights
- Licence terms per versionSource, 12 Nov 2026 (opens in a new tab)
- Hosted access
- Third-party host pricing: not yet checked
Where you can use it
- Download
- From Meta's Llama page, subject to the licenceOur draft
- Run locally
- With tools such as Ollama or llama.cpp (see Resources)Our draft
Before you choose
- Read the licence for the exact version you use. Terms differ between versions.
- Meta now also publishes its Muse models. They are a separate family and not covered here.
Put it to work
What changed
- 11 Nov 2026Re-checked Meta's Llama page; added Meta's description of Llama 4. No other changes.
Sources
- Meta: Llama models and licences (opens in a new tab)Checked 12 Nov 2026