Llama (Meta)

Llama is Meta’s family of open-weight LLMs (since Feb 2023) — the family that catalyzed the whole open-weight wave this spoke tracks. Provider: Meta AI. Source: Wikipedia.

Versions

• Llama 1 (Feb 2023, research-only weights) → Llama 2 (Jul 2023, broader release + instruct) → Llama 3 / 3.1 (8B/70B/405B) → Llama 4 (2025, mixture-of-experts, multimodal; Scout & Maverick). Lineup re-verified 2026-08-13 and unchanged — Llama 4 (5 April 2025) is still the latest, Scout 109B total / 10M context, Maverick 400B total / 1M context, both multimodal and multilingual across 12 languages; the 3.x line runs 3.0 (Apr 2024, 8B/70.6B, 8K ctx) → 3.1 (Jul 2024, adds 405B, 128K ctx) → 3.2 (Sep 2024, 1B/3B/11B/90B) → 3.3 (Dec 2024, 70B). Sixteen months with no new generation is itself the fact worth holding for a spoke that tracks release cadence — also a founding member of this spoke’s open-weight-models roundup.

The open-weight catalyst — and a licensing asterisk

The March 2023 leak of Llama 1 weights is widely credited as the “Stable Diffusion moment” for LLMs. But Llama is open-weight, not OSI-open: released under a custom community license with acceptable-use restrictions — the OSI disputes Meta’s “open source” claim (“do not meet … the Open Source Definition”). A concrete data point for the spoke’s licensing-clarity thread: contrast the Apache-2.0 cleanliness of gemma-4/mistral-ai with Llama’s source-available license.

Related

open-weight-models · llm-provider · gemma-4 · mistral-ai · deepseek · llm-benchmarks