Open Weights

Open weights refers to the practice of publicly releasing the trained parameter weights of an artificial intelligence model under a licence that permits users to download, inspect, modify, fine-tune, and deploy the model on their own hardware. Open-weight models — sometimes referred to loosely as "open-source AI" — have become a major force in the AI ecosystem, enabling researchers, startups, and independent developers to build on state-of-the-art foundation models without dependency on the API infrastructure or pricing of a single provider. The open-weight movement is led by organisations including Meta (Llama), Mistral AI, DeepSeek, Alibaba (Qwen), and Zhipu AI (GLM), and contrasts with the closed-weight approach of OpenAI, Anthropic, and Google's frontier models.^[1]

Distinction: Open Weights vs Open Source

A precise terminology distinction is increasingly used in the AI community. Open weights means the model's trained parameters are publicly available for download and use. Open source in the traditional software sense additionally requires the release of training data, training code, and reproducible training procedures. Very few large models meet the full open-source definition. The Open Source Initiative (OSI) published a formal Open Source AI Definition in October 2024 that distinguishes between these tiers, and most models commonly called "open source" are more accurately described as open-weight under that definition.

Key Open-Weight Model Families

Llama (Meta): Meta's Llama family is the most widely adopted open-weight model line. Llama 2 (July 2023), Llama 3 and 3.1 (2024), and subsequent releases have been released under the Llama Community Licence, which permits commercial use with restrictions for very large users. Llama models have been downloaded hundreds of millions of times and serve as the foundation for thousands of derivative models.

DeepSeek: Chinese AI company DeepSeek released DeepSeek-V2, V3, and the reasoning-focused DeepSeek-R1 as open-weight models under the MIT licence, making them among the most permissively licensed frontier-class models available. The release of DeepSeek-R1 in January 2025 — a reasoning model competitive with OpenAI's o1 — under an open licence was a landmark moment for the open-weight ecosystem.^[2]

Mistral AI: The French AI company Mistral has released several models (Mistral 7B, Mixtral 8x7B, Mixtral 8x22B) under Apache 2.0, one of the most permissive licences available.

Qwen (Alibaba): Alibaba's Qwen series covers a wide range of model sizes with strong multilingual capabilities, released under Apache 2.0 for most variants.

GLM (Zhipu AI): Zhipu's ChatGLM and GLM-4 models have been released with open weights, contributing to the Chinese open-weight ecosystem.

Advantages of Open Weights

Local deployment: Open-weight models can be downloaded and run on local hardware, eliminating per-query API costs and enabling use in offline, air-gapped, or latency-sensitive environments. Tools such as Ollama, vLLM, and llama.cpp make local deployment increasingly accessible.

Customisation and fine-tuning: Because the weights are available, users can fine-tune models on domain-specific data to produce specialised variants (medical, legal, financial) that outperform general-purpose API models on targeted tasks.

Data privacy: Running models locally eliminates the need to send sensitive data to external API providers, which is a significant advantage for regulated industries and government applications.

Transparency and research: Researchers can inspect model weights, study internal representations, and conduct safety research that is impossible with closed-weight models.

Limitations and Criticisms

Compute requirements: Running large open-weight models requires significant hardware. While quantisation techniques (reducing parameter precision) have made 7B to 70B parameter models runnable on consumer hardware, the largest models still require data centre-class infrastructure.

Training data opacity: Even when weights are open, the training data and procedures are typically not fully disclosed, making it difficult to audit for biases, copyrighted content, or safety issues embedded in the model.

Licence variability: Open-weight licences vary significantly in their terms. The Llama Community Licence imposes use restrictions based on user base size, and some "open" releases include acceptable-use policies that limit certain applications. Users must carefully review licence terms for each model.

Safety concerns: The release of powerful model weights has raised concerns about misuse — including for generating disinformation, deepfakes, or assisting with harmful activities — since open weights cannot be revoked once released and cannot be filtered at the point of use in the way that API-gated models can.^[3]

The Open-Weight Ecosystem

A rich ecosystem of tools and platforms has emerged around open-weight models:

• Hugging Face: The central hub for hosting, discovering, and distributing open-weight models, with over one million model repositories.
• Ollama: A tool for running LLMs locally with a single command, supporting Llama, Mistral, Qwen, and many other families.
• vLLM: A high-throughput inference engine optimised for serving open-weight models in production.
• llama.cpp: A C++ implementation enabling LLM inference on consumer hardware including CPUs and Apple silicon.
• LM Studio, GPT4All: Desktop applications that make open-weight models accessible to non-technical users.

• Fine-tuning
• Quantisation
• Ollama
• Hugging Face
• DeepSeek
• Llama

References