Open-weight model

An open-weight model is one where the trained parameters, the weights, are released publicly. Families such as Llama, Mistral, Qwen and Gemma fall into this group. You download the files, load them into an inference server and run them wherever you like. This is what makes self-hosting possible: without published weights there is nothing to host.

"Open-weight" is deliberately narrower than "open source". Open source implies you also get the training data and code, and a licence that allows any use. Most open-weight releases give you the weights and a licence with conditions, sometimes restricting commercial use above a user threshold or in certain fields. Reading the licence is part of the selection work, not an afterthought.

The open-weight ecosystem moves fast, and the gap to hosted frontier models narrows and widens by task. For extraction, classification and retrieval-grounded answering, mid-sized open-weight models are often sufficient. For long multi-step reasoning, hosted models still tend to lead. The honest answer comes from running your evals against both.

• CTO / Head of Engineering: open weights are the only route to running models inside the perimeter, and the licence dictates what you can legally ship.
• CFO: open-weight models remove per-token vendor pricing, turning inference into an infrastructure cost that you can forecast and depreciate.
• Compliance officer: a model you host can be version-pinned and audited; a hosted API can change under you without notice.
• Product manager: fine-tuning an open-weight model on your own data is possible in a way that closed APIs rarely allow.

Open-weight models exist because a single-vendor dependency for a core capability is a strategic risk. When a provider deprecates a model, raises prices or changes safety behaviour, an application built solely on it has no fallback. Published weights let organisations pin a version, run it privately and switch between models on their own timetable. The trade-off is that you take on serving, evaluation and upgrade work yourself, and you must read licences carefully. Used as part of a model-agnostic design, open weights give you leverage; used as a matter of ideology, they can leave you with a worse model than you need.

• A bank pinning a specific open-weight model version for credit-memo drafting so that behaviour is stable across an audit cycle.
• A SaaS vendor routing simple classification traffic to a small open-weight model and reserving hosted frontier models for complex reasoning.
• A healthcare provider fine-tuning an open-weight model on anonymised clinical vocabulary for better protocol retrieval.
• An education platform running a multilingual open-weight model for Hindi and Kannada student queries on its own GPUs.
• A logistics company deploying a small vision-language model on-site to read damaged shipping labels.

ConceptAn idea about how models are distributed and licensed, with practical consequences for hosting and cost. Eazyware evaluates open-weight and hosted models side by side on your tasks as part of Private Agentic AI and model selection work.

Eazyware service that covers it: Agentic AI Solutions (self-hosted). Starting prices are on the pricing page.