What Are Open-Weight AI Models? A Plain-English Guide
Open weight vs open source vs closed
The three labels describe different levels of openness. A closed model (like a commercial API) hides its weights completely - you use it remotely and cannot inspect it. An open-weight model publishes the weights but may keep training data, code and documentation proprietary. Truly open-source AI releases the weights, the code, and usually the data and benchmarks too. Most 'open' AI announcements today are actually open-weight.
How open-weight models work
Model weights are released as large files - often tens or hundreds of gigabytes. Anyone with a sufficiently powerful computer (or cloud rental) can load them using a framework like PyTorch, run inference locally, and fine-tune them on their own data with modest effort. This is the same technology powering chatbots, but running on hardware you control.
The trade-offs
For developers, open weights mean lower cost, full data privacy, customisation and no vendor lock-in. For society, they raise real concerns: unrestricted access can enable misuse, and most small teams lack the resources to fine-tune giant models safely. There is also a fairness argument - open-weight releases let researchers audit models rather than trusting vendor claims.
Why it matters now
Frontier open-weight models now rival the best closed systems on many benchmarks, and a broad coalition of companies and researchers is lobbying against restrictions (see our analysis of the the open-weights policy debate). Regulators are weighing security risks against innovation and sovereignty - expect this debate to define AI policy over the next two years.
The licence fine print
Open weight does not mean unrestricted. Meta's Llama models ship under a bespoke community licence with scale thresholds and branding conditions; Mistral releases many models under the permissive Apache 2.0; other families sit somewhere in between. The Open Source Initiative, which maintains the canonical definition of open source, published an Open Source AI Definition in 2024 arguing that true open source requires enough information about training data and code to rebuild the system - a bar most open-weight models do not meet. The distinction matters legally: if you deploy a model commercially, the licence, not the download link, determines what you may do.
Running one yourself
Open-weight models span roughly four orders of magnitude in size, from compact one-billion-parameter models that run on a phone to frontier-class models that need a rack of datacentre GPUs. Quantisation - storing weights in fewer bits - shrinks memory requirements about fourfold with a modest quality penalty, and tools such as Ollama and llama.cpp make it practical to run mid-size models on a gaming laptop. The usual motives are privacy (data never leaves the machine), offline reliability, freedom from per-token API pricing, and the ability to fine-tune behaviour for a niche task.
Where the ecosystem is heading
Three trends are reshaping the field. First, fine-tuning has become commodity work: adapting an open model to a niche domain is an afternoon job with a few hundred good examples, not a research project. Second, smaller models keep improving faster than their larger siblings, because labs distil frontier-model outputs into compact students - which makes the practical gap between a local open model and a hosted frontier API narrower than the raw benchmarks suggest. Third, agentic workloads, where software calls a model in a loop with tools, reward operators who control the full stack; open weights let teams optimise latency, cost and privacy together rather than accepting a hosted API's defaults.
Frequently asked questions
Are open-weight and open source the same?
Not quite. Open weight means the trained model file is public; open source usually implies the code, data and training recipe are public too. Many 'open' releases are open-weight only.
Are open-weight models safe?
They are a double-edged sword: transparency enables auditing and customisation, but unrestricted downloads also make misuse easier to replicate. Guardrails are typically built in but can be stripped by someone with technical skill.
What are examples of open-weight models?
Well-known families include Meta's Llama series and DeepSeek's models, which you can download and run locally or fine-tune. They power many self-hosted chatbots and enterprise deployments.
Are open-weight models safe to distribute?
Distributed weights can be fine-tuned to remove safety training, which is why licences add usage restrictions and why labs publish safety evaluations alongside releases. Regulation is catching up: the EU AI Act's rules for general-purpose AI treat very capable open-weight models differently from closed ones.
Why would a company give away model weights?
Reasons include ecosystem pull - developers build tooling around your stack - commoditising rivals' closed APIs, recruiting and brand value, and public or institutional mandates. Revenue then comes from services around the model rather than the weights themselves.
Sources & method: Hugging Face Docs · OpenAI Research · EFF. · Luminesca News publishes plain-English explainers built with AI-assisted drafting and a published source list. · Open Source Initiative · Ollama · Back to Luminesca News