Meta Ships Llama 4.5: A Mid-Cycle Open-Weight Refresh Reasserts Commitment to Open Source

Meta Ships Llama 4.5: A Mid-Cycle Open-Weight Refresh Reasserts Commitment to Open Source
📑 Table of Contents

Meta has shipped Llama 4.5, a mid-cycle open-weight refresh of the Llama 4 family, landing in early June 2026. The release updates both the Scout and Maverick variants with incremental capability gains, improved agentic tool-use stability, and modest license clarifications — all while keeping the broad community-license terms intact.

The timing is significant. After the April 2026 launch of Muse Spark — Meta Superintelligence Labs’ closed-weights reasoning model — there were genuine questions about whether Meta would continue investing in the open-weight Llama line. Llama 4.5 answers that question with a clear “yes.”

What’s New in Llama 4.5

The 4.5 refresh is not a generational leap but a targeted polish of the MoE (Mixture-of-Experts) architecture introduced with Llama 4 in April 2025. Scout and Maverick both benefit from:

  • Incremental benchmark gains across reasoning, coding, and multilingual tasks
  • Improved agentic tool-use stability — a meaningful upgrade for developers building autonomous AI workflows
  • License clarifications that ease enterprise adoption without changing the community-license framework
  • 512K context window support, keeping pace with the extended-context trend sweeping the frontier model landscape

The long-rumored Behemoth (2 trillion parameters) remains unreleased, leaving Maverick — with its 400B total / 17B active MoE configuration — as Meta’s open-weight flagship.

The June Frontier Wave

Llama 4.5 arrives amid an extraordinary density of model releases. June 2026 has seen a cascade of frontier and near-frontier models: DeepSeek V4.1, Qwen 3.7, GLM-6, and Gemma 4 have all hit HuggingFace in rapid succession. In this crowded field, Llama 4.5 is Meta’s bid to stay competitive in the open-weight tier while its proprietary Muse Spark pushes the frontier on the closed side.

The HuggingFace trending charts for June 2026 tell the story: Llama 4.5 sits alongside DeepSeek V4.1 and Qwen 3.7 among the most-downloaded models, with Chinese labs (Zhipu, Alibaba, DeepSeek, Moonshot) holding the top benchmark positions. Meta’s refresh ensures the Western open-weight ecosystem has a credible contender in an increasingly Asia-dominated landscape.

Open-Weight Commitment, Not Open-Source Purity

It’s worth noting the distinction: Meta’s models are open-weight, not open-source in the traditional sense. The Llama 4 Community License imposes usage restrictions that the OSI would not recognize as open-source. But for the vast majority of developers and enterprises, downloadable weights with permissive deployment terms are what matters — and Llama 4.5 delivers on that promise.

With 1 billion Llama downloads celebrated in March 2025 and a sprawling ecosystem of fine-tunes, quantizations, and inference engines (llama.cpp, vLLM, Ollama), the Llama brand remains the infrastructure layer of the open-weight world. Llama 4.5 ensures it stays that way — at least until Llama 5 arrives.