Meta releases Llama 4.1 70B Thinking — open-weight reasoning model lands with day-zero llama.cpp and vLLM support, MIT-aligned community license updated
Meta released Llama 4.1 70B Thinking on May 28 — an open-weight reasoning model that lands with day-zero support across llama.cpp, vLLM, and the major inference-runtime ecosystem. The community license was updated to align more closely with MIT-style permissive terms while retaining the acceptable-use restrictions, and the model is positioned against Mistral Large 3 and the broader open-thinking-tier emergence.
The model release is the substantive piece. Llama 4 — released in early 2026 — established Meta's open-weight frontier with the 70B and 405B variants but did not include an explicit reasoning-mode-trained checkpoint. Llama 4.1 70B Thinking adds the reasoning-mode-trained variant, with the model trained on a combination of supervised reasoning traces and Scale AI's RLHF pipeline. The day-zero inference-runtime support — llama.cpp, vLLM, TGI, and the major commercial inference-providers (Together, Fireworks, OctoAI) all ship support on release day — is the operational shape of Meta's open-source strategy: ship the weights with the ecosystem already aligned.
The competitive context is the open-thinking-tier emergence. Mistral Large 3 launched the same day under Apache 2.0, also with a reasoning-mode variant. The two releases mark the first time the open-weight ecosystem has shipped competitive reasoning-mode-trained models alongside the closed-frontier-lab releases (GPT-5.2 Thinking Mode, Claude Opus 4.7). For enterprises evaluating self-hosted-versus-API deployment, the May 28 release window collapses the capability-gap argument that previously favored the closed-frontier-lab tier for reasoning-heavy workloads. The license-update side — Meta's move toward MIT-aligned permissive terms — addresses the commercial-deployment friction that the prior Llama community license created for some enterprise integrators.
Meta AI — Llama 4.1 70B Thinking open-weight release May 28 2026 → · Hugging Face — Llama 4.1 70B Thinking model card and weights → · arXiv — Llama 4.1 reasoning-mode training methodology →