Llama

A family of open-source large language models from Meta, and one of the most active open-source foundation models in the global community ecosystem.

Llama is a family of open-source large language models released by Meta (Facebook’s parent company), and one of the most active foundation models in today’s global open-source community. Since Llama 1 was released under a limited research-only license in 2023 and Llama 2 relaxed its commercial licensing later that year, Llama has quickly become the default choice for countless AI applications, fine-tuned models, and private deployment solutions. In 2024, Llama 3.1 pushed the flagship model to 405B parameters and extended the context window to 128K; in 2025, Llama 4 delivered major upgrades in multimodal understanding, reasoning and planning, and Agent capabilities.

Pricing: completely free and open source. Meta publishes the Llama model weights publicly, so developers can download and run them locally, on private clouds, or on edge devices. Llama 3.x uses the Llama 3 Community License, which is highly friendly to most commercial applications, but large platforms with more than 700 million monthly active users must apply to Meta for a special license. Cloud providers (AWS, Azure, Google Cloud, Alibaba Cloud, etc.) also offer managed APIs billed by usage.

In terms of ecosystem positioning, Llama is the “default option” among open-source large models: the most complete tooling ecosystem, the largest collection of community fine-tuned models, and the richest enterprise adoption cases. Compared with Qwen, Llama offers stronger English capabilities and more international community resources; compared with closed-source models like GPT-4 and Claude, Llama is better suited for scenarios where data must stay in-house, cost control is strict, or deep customization is required. If you need to build an LLM foundation that supports private deployment and long-term iteration, Llama is a candidate you simply can’t overlook.

Note: While Llama’s Chinese capabilities have become quite good through community fine-tuning, the native versions still lag behind Chinese-optimized models like Qwen; the large-parameter versions (70B/405B) require professional GPU clusters, so make sure to evaluate your hardware and quantization plan before deploying locally.

Model Iteration Timeline

Time	Version/Model	Highlights
2023-02	Llama 1	Meta releases LLM weights publicly for the first time (research use)
2023-07	Llama 2	Relaxed commercial licensing, three sizes: 7B/13B/70B
2024-04	Llama 3	New-generation training data and architecture, 8B/70B debut
2024-07	Llama 3.1	405B flagship released, 128K context
2025-04	Llama 4	Upgrades in multimodal, reasoning, and Agent capabilities

Source: Agentopia · Site-wide model catalog at /llms.txt