Meta has released Llama 2 for research and commercial use, giving developers downloadable language-model weights and a choice of cloud deployment routes. Microsoft is its preferred partner, with the models available in the Azure AI catalogue.
Meta’s headquarters in 2022, not the Llama 2 launch. Photo: LPS.1, “Meta Platforms Headquarters Menlo Park California” (2022). CC0 1.0. Resized; cropped; responsive crops.
The launch on 18 July includes pretrained models and Llama 2-Chat, versions tuned for conversation. Each comes in three sizes: 7 billion, 13 billion and 70 billion parameters.
Weights, cloud deployment and Windows
Meta’s announcement includes the weights and starting code. Its release instructions require applicants to accept the licence and obtain approval before receiving an emailed download link. Hugging Face access also requires a request using the same email address as the Meta application.
Microsoft says Azure customers can fine-tune and deploy all three model sizes. Its model catalogue is in public preview, with tools for evaluation and deployment and Azure AI Content Safety integrated into the deployments. The company also outlines local Windows support through ONNX Runtime and DirectML, describing that optimisation in future tense.
Microsoft’s preferred-partner role doesn’t make Azure the sole route. Meta also names AWS and Hugging Face.
Downloading weights lets developers operate a model themselves; deploying through a cloud platform uses that provider’s infrastructure. A royalty-free model licence doesn’t supply the computing resources needed to run it.
What the commercial licence permits
The July 18 licence permits use, modification and redistribution subject to its terms. Distributors must pass along the agreement and keep its attribution notice in a Notice text file.
The separate-permission rule for organisations above 700 million monthly active users looks at June 2023 usage, assessed at release. It covers products or services offered by or for the licensee or its affiliates, not just a Llama-powered app. Meta must grant permission before those organisations exercise the licence’s rights. Section 2 doesn’t apply to recipients who receive the materials as part of an integrated end-user product.
Another condition matters for model builders: the agreement bars using Llama materials, outputs or results to improve another large language model, except Llama 2 or its derivatives. Downloadable weights don’t bring an unrestricted right to use their output for training other LLMs.
The licence also incorporates an acceptable-use policy. Among its provisions are restrictions on malicious code and deceptive uses, including presenting outputs as human-generated. Developers must appropriately disclose known dangers of their AI systems to end users.
Meta describes the release as open source. The concrete offer is downloadable weights under a custom commercial licence with conditions.
What changed inside the models
The launch model card specifies a 4,096-token context window and two trillion pretraining tokens. Meta says neither the pretraining nor the fine-tuning data includes its users’ data. The pretraining cutoff is September 2022, although some tuning data extends to July 2023. These are text-input, text-output models, with commercial and research use in English as the stated scope.
Llama 2-Chat adds supervised fine-tuning and reinforcement learning from human feedback. The pretrained models remain a separate starting point for developers adapting text-generation systems. Among the released sizes, the 70B model uses grouped-query attention to improve inference scalability.
The first version of the technical paper also reports a 34B model, but that size isn’t part of the release. The authors say they’re delaying it because they haven’t had enough time for sufficient red-teaming. Benchmark tables that include 34B therefore describe a broader research set than the models developers can download.
Meta’s model card warns that outputs can be inaccurate, biased or objectionable and calls for safety testing and tuning specific to each application.
Earlier coverage: LLaMA.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
