Meta Llama ecosystem
The most-deployed open weights in the West, now in managed decline — Meta's frontier work moved to proprietary Muse, the Llama API is dead, and the license was never open source.
Element scores
Strengths
Free, battle-tested weights with the largest installed base in the West and the deepest community stack: llama.cpp and vLLM for inference, torchtune/Unsloth for fine-tuning, and the strongest open safety tooling anywhere (Llama Guard 4, LlamaFirewall, Prompt Guard 2). Llama 4 Scout's 10M-token context is class-leading and runs on a single H100 quantized. Self-hosting economics remain compelling for existing deployments — nothing gets rug-pulled that you already run.
Honest dings
The vendor has left the building: Meta's frontier development moved to the proprietary Muse line (April 2026), the Llama API shut down on 06/07/2026, and Llama Stack was handed to the community as the model-agnostic OGX. Llama 4 itself underperformed (benchmark-gaming admissions, developer traction below expectations) and Chinese open-weight rivals now outpace it. The license was never open source per the OSI — 700M-MAU gate, naming and attribution duties, a unilaterally updatable Acceptable Use Policy, and (critical for European readers) Llama 4 multimodal rights are explicitly denied to companies with their principal place of business in the EU. No official word on any future Llama release.
Sources
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- llama.developer.meta.com — Official: Llama API public preview retired 06/07/2026; users directed to third-party hosts; weights remain downloadable (accessed 2026-07-08)
- ai.meta.com/blog/Llama-4-multimodal-intelligence — Official Llama 4 announcement: Scout (17B active/109B, 16 experts) and Maverick (17B active/400B, 128 experts), natively multimodal MoE; Behemoth previewed but not released (accessed 2026-07-08)
- llama.meta.com/docs/deployment/versioning — Official deployment docs: Scout 10M-token / Maverick 1M-token context windows, Llama 3.x line specs, versioning and migration guidance (accessed 2026-07-08)
- about.fb.com/news/2026/04/introducing-muse-spark-me… — Official: Muse Spark (08/04/2026) is proprietary, powers Meta AI, private API preview only; 'hope to open-source future versions' — no commitment (accessed 2026-07-08)
- venturebeat.com/technology/goodbye-llama-meta-launc… — Independent: Llama 4 mixed reviews + benchmark-gaming admissions, 1.2B ecosystem downloads, Meta non-answer on future Llama development, Chinese rivals at 41% of HF downloads (accessed 2026-07-08)
- thenewstack.io/meta-abandons-llama-spark — Independent analysis (30/04/2026): Meta has practically abandoned Llama development; migration options for Llama users; Andrew Ng on the loss to the developer community (accessed 2026-07-08)
- llama.com/llama4/license — Llama 4 Community License: 700M-MAU clause requiring Meta permission, Acceptable Use Policy incorporated by reference (accessed 2026-07-08)
- opensource.org/blog/metas-llama-license-is-still-no… — Open Source Initiative: Llama licenses fail the Open Source Definition (freedom 0, OSD 5, OSD 6); Meta's 'open source' branding is open-washing (accessed 2026-07-08)
- compliance-made-simple.ch/llama — Independent compliance analysis: Llama 4 AUP denies multimodal-model rights to EU-domiciled individuals/companies; 'Built with Llama' attribution and 'Llama' naming requirements (accessed 2026-07-08)
- ai.meta.com/blog/llamacon-llama-news — Official LlamaCon (04/2025): Llama Guard 4, LlamaFirewall, Prompt Guard 2, CyberSecEval 4, 1B+ downloads, Llama API preview launch (since retired) (accessed 2026-07-08)
- github.com/meta-llama/llama-stack — Repo now redirects to community-run OGX (ex-Llama Stack): renamed, model-agnostic, MIT — evidence Meta's first-party harness was spun off (accessed 2026-07-08)