Meta Llama
Open SourceThe open-weight model family powering the AI ecosystem.
Scores
About
Meta Llama is Meta's family of open-weight large language models, representing the dominant force in the open-source AI ecosystem. Unlike proprietary models, Llama weights are publicly released under the Llama Community License, allowing developers and organisations to download, self-host, fine-tune, and deploy models without per-token cloud costs.
Llama 3 series (8B / 70B / 405B) Llama 3.1 introduced the 405B parameter model — the first frontier-level open-weight model competitive with GPT-4 class systems — alongside upgraded 8B and 70B variants. All Llama 3.1+ models support a 128K token context window, multilingual output across eight languages (English, German, French, Italian, Portuguese, Hindi, Spanish, Thai), and strong tool-use / function-calling capabilities. Llama 3.2 added vision: the 11B and 90B multimodal variants can reason over high-resolution images, while lightweight 1B and 3B text-only models target edge and on-device deployments. Llama 3.3 delivered a refined 70B instruction-tuned release with improved benchmark performance.
Llama 4 (Scout / Maverick / Behemoth preview) Llama 4, released in early 2026, marked a fundamental architectural shift to mixture-of-experts (MoE) and native multimodality. Scout has 17B active parameters across 16 experts (109B total) and an industry-leading 10 million token context window — fitting on a single H100 with INT4 quantisation. Maverick has 17B active parameters across 128 experts (400B total) and a 1 million token context window, delivering GPT-4o-class performance. Behemoth (in training / preview) targets ~2 trillion total parameters and is expected later in 2026.
Access methods
- Self-hosting: Download weights from Hugging Face (meta-llama organisation) or Meta's website. Run locally via Ollama (one-command setup), llama.cpp (CPU/GPU inference on consumer hardware), vLLM (high-throughput production serving), or Hugging Face Transformers.
- Managed APIs: Dozens of providers offer Llama via pay-per-token APIs — Groq (fastest raw throughput via LPU chips), Together AI, Fireworks AI, Deepinfra, Replicate, and major cloud platforms: Amazon Bedrock, Microsoft Azure AI, Google Cloud Vertex AI, and Cloudflare Workers AI.
- Meta AI product: Meta AI is the consumer chat interface available at meta.ai and embedded in WhatsApp, Instagram, and Facebook, powered by Llama models.
License Llama uses the Llama Community License — not an OSI-approved open-source license. Commercial use is permitted for most organisations. Key restrictions: if a product exceeds 700 million monthly active users, a separate agreement with Meta is required; Llama outputs cannot be used to train other foundation models; Llama 3.2 multimodal weights are not licensed for use by companies headquartered in the EU.
As of May 2026, the meta-llama/llama GitHub repository has approximately 59K stars, and the Llama model family on Hugging Face is among the most downloaded of any model family.
Key Features
- Multiple model sizes: 1B, 3B, 8B, 11B, 70B, 90B, 405B — covering edge to frontier use cases
- Llama 4 Scout: 10 million token context window with MoE architecture (109B total params)
- Llama 4 Maverick: 1 million token context, 400B total parameters, GPT-4o-class performance
- Native multimodality in Llama 3.2 (11B/90B vision) and the full Llama 4 generation
- Self-host via Ollama, llama.cpp, vLLM, or Hugging Face Transformers
- Available on Amazon Bedrock, Azure AI, Google Vertex AI, Groq, Together AI, Fireworks AI, and more
- Multilingual support across 8+ languages in instruction-tuned variants
- Fine-tuning friendly — standard transformer architecture with published training details
Pros
- Zero per-token cost when self-hosting — economics scale favourably for high-volume workloads
- Maximum portability: run on a laptop with Ollama, a server with vLLM, or any managed cloud API
- Largest open-weight model ecosystem — more tooling, fine-tunes, and community support than any other family
- Llama 4 Scout's 10M context window is unmatched in any open-weight model
- Fine-tunable on domain-specific data without vendor permission or API access
- Competitive benchmark performance — Llama 3.1 405B and Llama 4 Maverick rival proprietary frontier models
- Available on virtually every cloud and inference provider — no single-vendor dependency
Cons
- Self-hosting requires GPU hardware, infrastructure knowledge, and ongoing maintenance
- The Llama Community License is not true open source — restricts training other models and large commercial deployments (>700M MAU)
- Llama 4 Maverick/Scout require significant GPU memory for full-precision self-hosting
- Managed API providers charge per token, removing the cost advantage of self-hosting
- Multimodal Llama 3.2 weights are not licensed for EU-headquartered companies
- Quality gap vs. frontier proprietary models (GPT-4.5, Claude 3.7) for the most complex reasoning tasks
Pricing
Open SourceRelated Tools
Alternative to (8)
Learning Resources
No resources yet — check back soon.
