Llama

TAS Score: S3/3 — D4/5 — A3 / T1 (T1 not T2: the Llama Community License is not OSI-approved — it adds a 700M-MAU commercial threshold, a competitor restriction, a ban on using Llama to train other models, and (since 3.2) an EU multimodal exclusion. Weights are downloadable and self-hostable, so autonomy stays A3; license transparency is capped at T1.) (D4 not D5: Meta grants the >700M-MAU license at its “sole discretion,” and the Acceptable Use Policy is incorporated by reference — a standing conditionality on continued use at scale.)

Note: Meta publishes open weights (Llama 3.x / 4) but has signalled a pivot toward closed models (“Muse Spark”). This card covers the open-weight Llama family.

Brief Description

Meta’s open-weight model family — Llama 3.1 (8B/70B/405B), 3.2 (small on-device + multimodal), 3.3 (70B), and Llama 4 (Scout, Maverick — MoE, natively multimodal, released April 2025). Weights are downloadable from HuggingFace and self-hostable via Ollama, vLLM, llama.cpp, SGLang.

The catch is the license. Every Llama release ships under the Llama Community License, a bespoke commercial license the OSI has explicitly found fails the Open Source Definition: it caps free commercial use at 700M monthly active users, bans building competing models, prohibits using Llama or its outputs to train other models, and (from 3.2) excludes EU-domiciled parties from the multimodal models. Fully runnable offline, but not OSI-open — hence T1.

Architectural Role

Compute/inference layer: one of the most widely deployed open-weight bases, with a large ecosystem of derivatives and fine-tunes. Local alternative to cloud chat / multimodal APIs across the full size range.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally
  • Does not require external accounts (self-hosted)
  • Allows data export — weights downloadable (Llama Community License; MAU / competitor / training restrictions apply)
  • Commercial use conditional on Meta’s terms; EU multimodal excluded

Philosophical Assessment (whose.world criteria)

Criterion Status Comments
Pause ✅ Stop inference. Model weights stay on disk.
Exit ✅ Standard model format, downloadable.
Recoverability ✅ Re-download from HuggingFace or restore from backup.
Visibility ⚠️ Open weights, but non-OSI Community License (MAU cap, competitor & training bans, EU exclusion) → T1. Training data not published.
External Dependencies ⚠️ Self-hosted: none at runtime. License: conditional on Meta’s terms at scale.

Configuration (Minimal) 

# Consumer hardware via Ollama
ollama run llama3.3      # 70B
ollama run llama3.2      # small / multimodal

# Server — Llama 4 via vLLM
vllm serve meta-llama/Llama-4-Scout-17B-16E-Instruct --tensor-parallel-size 8

Alternatives

Alternative Autonomy Notes
Qwen3 A3 / T2 Apache open weights — no MAU cap, no competitor ban.
Gemma 4 A3 / T2 Apache — Google’s OSI-clean open family.
DeepSeek-R1-Distill-Llama A3 / T1 Llama-based distill — inherits the same T1.

Trajectory

Direction: closing

Llama popularized open weights, but the license has tightened over time (the EU multimodal exclusion arrived in 3.2), and Meta has signalled a strategic pivot toward closed “Muse Spark” models, casting doubt on long-term open releases. The weights you already hold stay usable, but the direction of travel is toward more restriction, not less — the mirror image of Gemma.

Signal assessment:

Signal Status Evidence
License ⚠️ Llama Community License — non-OSI; MAU cap, competitor & training bans.
Feature gating ⚠️ EU-domiciled parties excluded from multimodal models (since 3.2).
Self-hosting ✅ Standard formats. Ollama, vLLM, llama.cpp, SGLang.
Governance ⚠️ Meta controls terms; reported pivot toward closed models.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

This site uses Just the Docs, a documentation theme for Jekyll.