AI tools·Chat models·Global
Llama
Meta's open-weight model family with the world's largest self-hosting and fine-tuning ecosystem.
Compiled through 2026-09
Official app blocked; weights are open for self-hosting
- Vendor
- Meta
- Started
- 2023
- China availability
- Overseas accessRequires overseas network access
- Pricing tier
- Free / open
- Current flagship
- Llama 4.2 (Behemoth / Scout / Maverick)
- Website
- www.llama.com
Evaluation, Pros & Cons
Editorial score based on public benchmarks and community feedback — not a third-party benchmark result
Key Advantages
- Flagship of the open-weight ecosystem, with the richest fine-tuning and on-prem tooling
- Long-context MoE architecture scales from single GPU to clusters at low inference cost
Limitations & Caveats
- The official chat app is unavailable in China; requires self-hosting or third-party instances
- Chinese pretraining share is lower than domestic rivals; pick tuned variants for Chinese workloads
Overview
Since Feb 2023, the Llama family has carried the open-weight banner: from Llama 2's commercially friendly license, through Llama 3 closing in on closed flagships, to Llama 4's long-context MoE design. Spanning on-device quantized models to datacenter-scale experts, Llama remains the default base for self-hosted and fine-tuned AI, with tens of thousands of derivatives.
Direction
Meta is working on releasing a Behemoth-class frontier model, on-device small variants, and faster inference with cloud and OSS runtime partners.
Version history
Jun 2026
LatestLlama 4.2
New long-context MoE with stronger efficiency and tool use.
Apr 2025
Llama 4
MoE era begins with Scout and Maverick.
Jul 2024
Llama 3.1 / 3.2
128K context plus on-device 1B/3B models.
Apr 2024
Llama 3
8B/70B tiers approaching closed flagships.
Jul 2023
Llama 2
Commercial license unlocked an ecosystem boom.
Feb 2023
Llama 1
Research preview that kicked off the open-weight wave.