Llama is Meta's open-weights AI model family, and the single most important release program in open-source AI. Since the original Llama leaked in early 2023, Meta has shipped increasingly capable models under a near-open license that permits commercial use — powering an entire ecosystem of downstream AI products, research, and self-hosted deployments.
This page is the living reference to Llama — the model family, licensing, pricing, what's new, how it compares, and every story we publish tagged Llama. We update it whenever Meta ships a meaningful change.
As of April 2026, Llama 4 is the current generation — a mixture-of-experts family with Scout, Maverick, and Behemoth variants — and the frontier for open-weights AI.
We rated Llama 3.9 / 5. Llama 4 is still the default open-weights choice for anyone who needs to own their inference: permissive enough for most commercial use, cheap to rent, and hosted on every major cloud. The caveat is direction of travel — Meta has not published a new open generation since April 2025, and its 2026 frontier models ship behind a closed API.
Read the full Llama review →Latest news about Llama
Every story we've published tagged Llama. Refreshed automatically every time the newsroom publishes.
- Jun 28, 2026Google restricts Meta's access to Gemini AI modelsThe curb signals Google treating Meta as a direct AI rival rather than a customer, per a Financial Times report.
- May 5, 2026Macmillan, McGraw Hill, and Hachette sue Meta over Llama training dataFive publishers and author Scott Turow allege Meta pulled copyrighted books from LibGen and Anna's Archive to train Llama.
Timeline of major releases
- Feb 24, 2023Llama 1 leaksMeta releases Llama 1 to academic researchers under a restrictive license; weights leak on 4chan days later, kicking off the modern open-weights movement.
- Jul 18, 2023Llama 2 ships with commercial useMeta releases Llama 2 under a near-open license that permits commercial use. Sizes from 7B to 70B parameters. Becomes the default open foundation model.
- Apr 18, 2024Llama 3 launchesLlama 3 ships at 8B and 70B sizes. Meta follows with Llama 3.1 (including a 405B variant) in July, trading blows with GPT-4-class closed models on benchmarks.
- Sep 25, 2024Llama 3.2 adds visionMeta adds multimodal variants (11B and 90B) and small on-device models (1B and 3B). Llama becomes viable for edge deployment.
- Apr 5, 2025Llama 4 releasedMeta ships Llama 4 as a mixture-of-experts family — Scout (small), Maverick (mid), and Behemoth (large). Native multimodal, 10M context window on Scout.
- Feb 28, 2026Llama 4.1 and the research arm restructuringMeta reorganizes its AI research into a unified "Superintelligence Labs" org. Llama 4.1 ships with improved reasoning; Meta doubles down on the open-weights strategy.
How to access Llama
| Plan | Price | Includes |
|---|---|---|
| Llama weights | Free | Download weights for Llama 4 Scout, Maverick, and prior releases. Subject to Meta's Llama Community License. |
| Meta AI (consumer) | Free | Llama-powered assistant inside WhatsApp, Messenger, Instagram, and on the web. |
| Llama API (preview) | Usage-based | Meta's own hosted Llama API, launched in 2025. Pricing competitive with OpenAI Flash-class models. |
| Third-party hosted | Usage-based | Llama available on AWS Bedrock, Azure AI, Google Cloud Vertex, Groq, Together, Fireworks, and many other inference providers. |
Llama vs. ChatGPT vs. Claude
| Feature | Llama | ChatGPT | Claude |
|---|---|---|---|
| License model | Open weights (Llama Community License) | Closed / hosted only | Closed / hosted only |
| Self-hostable | Yes | No | No |
| Largest context window | 10M (Llama 4 Scout) | 200K (GPT-5 Pro) | 1M (API) |
| Multimodal | Yes (native) | Yes | Images only |
| Price per token | Near zero (self-host) / very low (providers) | Higher | Higher |
| Benchmark frontier | Trails GPT-5 / Claude Opus by a generation | Frontier | Frontier |
The Llama family and MoE architecture
Llama 4 shifted the family to mixture-of-experts. Instead of one dense model per size, each Llama 4 model routes each token through a small subset of expert networks — so the compute cost is much lower than the parameter count suggests. Scout is positioned for efficient inference and ultra-long context (10M tokens). Maverick is the mid-size default for most production workloads. Behemoth is the flagship, used to distill and improve the smaller models.
This architecture is the same bet OpenAI and DeepSeek have made with their frontier models, and it's now the default for cost-efficient scaling in frontier AI.
The open-weights strategy
Meta's open-weights thesis is straightforward: by releasing strong weights publicly, Meta commoditizes the model layer while capturing most of the distribution layer (billions of Meta app users with built-in AI assistants). It's a distribution-first strategy that makes closed-model pricing harder to sustain over time.
For developers and enterprises, this is the gift. Llama weights downloaded today run forever, on your hardware, with no per-token cost and no vendor lock-in. For many production AI workloads — especially high-volume, latency-sensitive, or privacy-sensitive ones — that's the decisive advantage over closed APIs.
How Llama is used in 2026
Most Llama usage isn't at llama.meta.com — it's inside downstream products. Meta AI (the assistant inside WhatsApp, Instagram, Messenger) is Llama-powered. So are many third-party AI products that self-host Llama variants. Cloud providers — AWS Bedrock, Azure, Vertex, plus specialty inference providers like Groq and Together — have made Llama the default open foundation model they offer.
For teams deciding whether to use Llama, the practical question is: do you need frontier capability, or do you need enough capability with control? Frontier reasoning tasks (novel proofs, long multi-step agents, advanced coding) still tilt toward Claude or GPT-5. High-volume, steady-state workloads (RAG, classification, summarization, content moderation) tilt toward Llama.
Where Llama wins — and where it doesn't
Llama wins on three axes closed models can't match: cost at scale (nearly free at self-hosted volume), deployability (on-prem, air-gapped, edge), and weight ownership (you control the model, not a vendor). For teams that care about any of those, it's the default.
It trails the frontier on the hardest reasoning tasks, and the ecosystem around tool use, coding agents, and multimodal polish is still catching up to OpenAI and Anthropic. The gap has narrowed each generation but hasn't closed.
Llama license and commercial use in 2026
Llama is "open-weights" rather than fully open-source — the weights and inference code are released, but the license is Meta's own Llama Community License, not Apache or MIT. The terms have stayed consistent across Llama 3 and Llama 4: free for research and commercial use, with one significant restriction. Companies with more than 700 million monthly active users at the time of the model's release must request a separate license from Meta. That clause is aimed at Apple, Google, ByteDance, and Tencent — not at the long tail of developers and enterprises that make up most Llama users.
The license also includes a use-policy attachment that prohibits the usual list (CSAM, weapons, election interference) and an attribution requirement: products built on Llama must acknowledge the model in user-facing documentation. There is no royalty, no per-token fee, and no obligation to send traffic data back to Meta.
For most production deployments, the practical takeaway is: yes, you can use Llama 4 commercially, you do not pay Meta, and you should keep an "About this AI" note in your product docs that mentions the model lineage. If you're at the 700M-MAU scale where the license trips, you should already be talking to Meta directly.
Running Llama: self-host, inference providers, and on-device
Most teams don't run Llama on their own GPUs. They use one of the inference providers that has built a business around hosting open-weights models — Together, Fireworks, Groq, Hyperbolic, Replicate, Modal, AWS Bedrock, Azure AI Foundry, and Vertex Model Garden all offer Llama 4 endpoints with OpenAI-compatible APIs. Pricing varies meaningfully (often 3-5x between cheapest and most expensive on the same model), and Groq remains the latency leader thanks to its custom inference chips.
Self-hosting is the right move when data residency, custom fine-tuning, or per-token economics at high volume justify the operational overhead. Llama 4 Scout fits on a single H100 (or two consumer GPUs for quantized inference); Maverick wants 4-8 H100s; Behemoth is a multi-node deployment.
On-device, Llama 3.2 1B and 3B remain the practical defaults for phones and laptops via llama.cpp, MLX, and Ollama. Llama 4's MoE architecture doesn't currently target edge devices the way the dense Llama 3 small models did — for now, on-device Llama and frontier Llama are different families.
Our coverage
Everything tagged Llama appears in Latest news above. For Meta's broader AI strategy — Reality Labs, Meta AI assistants, the Superintelligence Labs restructure — see our reporting under the AI news and AI business beats. For other open-weights models, see DeepSeek and Mistral.
