Open-Weight Models

Open-weight models — often loosely called open-source AI models — are artificial intelligence models whose trained parameters, or weights, are released publicly so that anyone can download, run, fine-tune or redistribute them. Unlike closed models, which users reach only through a vendor's API, an open-weight model can be deployed on local or private infrastructure, giving organisations direct control over their data, costs and customisation. Strictly speaking, the label describes what is released: the weights. The training data, training code and recipes may remain private, which is why researchers distinguish open weights from fully open-source AI.^[1]

Background

The category entered the mainstream when Meta released its Llama model family under permissive licences, followed by European laboratory Mistral and a wave of Chinese developers — DeepSeek, Alibaba (Qwen), Zhipu AI (GLM), Moonshot AI (Kimi) and MiniMax — whose releases regularly topped open leaderboards. By mid-2026, open-weight models had closed much of the performance gap with proprietary systems. In OpenRouter's June 2026 survey of open models, Zhipu's GLM 5.2 ranked first among open-weight models on an independent intelligence index at a score of 51, ahead of NVIDIA's Nemotron 3 Ultra at 48 and MiniMax M3 and DeepSeek V4 Pro at 44, and within roughly five points of the leading closed model at the time. NVIDIA's Nemotron line, built for enterprise deployment, became the strongest United States open-weight entrant.^[1][6]

How They Are Used

Open weights change both the economics and the governance of AI. Organisations can run models on their own hardware, keeping sensitive data inside their networks and avoiding per-token fees; they can fine-tune a model on domain data, distill smaller models for edge devices, or quantize models to run on cheaper hardware, supported by a growing ecosystem of inference servers and deployment platforms. Licences vary widely: some releases use permissive terms such as MIT or Apache, others use bespoke community licences with usage restrictions, and at least one 2026 frontier release had its weights held for a two-week safety review before publication — a sign of how safety evaluations and open distribution are being reconciled in practice.^[2][4]

Adoption data reflects the shift. Vercel, which routes production AI traffic for thousands of applications, reported in June 2026 that open-weight models handled 29 per cent of the tokens passing through its gateway, up from about one-ninth two months earlier. Enterprise interest centres on data control and cost, and governments cite open weights as a foundation for sovereign AI programmes that must operate within national jurisdictions and legal systems.^[2][5][6]

Significance

Open-weight releases exert competitive pressure on closed-model pricing, spread capability faster than any single vendor can, and allow smaller countries and companies to participate in AI development without building frontier models from scratch. They also raise policy questions — about safety testing, misuse and whether advanced AI capability should be freely downloadable — that regulators and standards bodies in the United States, Europe and Asia continue to debate. Major technology firms have taken public positions on both sides, with some arguing that open weights strengthen domestic AI ecosystems and competition, and others emphasising security and control.^[3][5]

>See Also

• Sovereign AI
• Llama
• DeepSeek
• Qwen
• Zhipu AI
• Kimi AI
• Fine-Tuning
• Quantization
• YTL AI Cloud