# Open-weight model

A model whose trained parameter weights are available for download and use under stated license terms, enabling inspection, adaptation, or deployment outside a hosted API.

technical-term · Models & inference · foundational · Reviewed 2026-08-09

## Definition

An open-weight model makes trained parameters available under a specific license and usage policy. That can enable local inference, self-hosting, quantization, adaptation, and independent evaluation, but “open weight” does not by itself mean open source, unrestricted use, reproducible training, disclosed data, or safe deployment. The exact artifact, license, model card, and runtime dependencies determine what is actually possible.

Weight distribution is established; terminology, licenses, training transparency, and obligations vary substantially across model families.

## Why it matters

Weights are only one layer of the system. Tokenizers, inference code, adapters, safety components, data provenance, and acceptable-use terms can carry separate conditions.

Availability changes who can operate and modify a model, which increases deployment choice and also transfers more evaluation, security, update, and incident responsibility to the operator.

## System anatomy

- **Weight artifact:** The trained parameters and supported numerical formats.
- **License and policy:** Permissions, restrictions, attribution, and downstream obligations.
- **Model documentation:** Architecture, evaluations, intended uses, and known limitations.
- **Runtime compatibility:** Software and hardware required to execute the artifact.



## Important distinctions

- **Open-source AI system:** Open weights may omit training code, data, or reproducibility information and should not be labeled open source without meeting the relevant definition.
- **Free hosted model:** Free access to an endpoint does not provide the model weights or deployment rights.



## Implementation signals

- Review the exact artifact license and usage policy
- Record model, tokenizer, quantization, adapter, and runtime versions
- Run workload-specific quality and safety evaluation before deployment

## Failure modes

- A team assumes open weights permit every commercial or regulated use
- A repackaged checkpoint loses provenance and model-card context
- Downloadability is mistaken for privacy or safety



## Related knowledge

- [Local inference](https://aisdlc.ai/agentic-engineering/local-inference) — Model inference executed on the user’s or organization’s local hardware and runtime rather than delegated to a separately operated remote model endpoint.
- [Self-hosted inference](https://aisdlc.ai/agentic-engineering/self-hosted-inference) — An inference service whose runtime, capacity, configuration, model artifacts, and operational controls are managed by the organization using it or by a directly governed operator.
- [Model adaptation](https://aisdlc.ai/agentic-engineering/model-adaptation) — The deliberate modification of a model or its task-facing behavior for a target domain, task, language, policy, or operating environment using measured evidence.
- [Quantization](https://aisdlc.ai/agentic-engineering/quantization) — The representation or computation of model values at lower numerical precision to reduce memory, storage, bandwidth, or inference cost, with workload-dependent tradeoffs.

## Sources and further study

- [OpenAI — Introducing gpt-oss](https://openai.com/index/introducing-gpt-oss/) — Introduces two Apache 2.0 open-weight reasoning models intended for customizable deployment, including consumer-hardware and local-inference scenarios. **Use:** First-party case study. **Limitation:** This is a first-party launch description. Benchmark claims do not establish local workload fitness, privacy, safety, governance, or operational readiness for a specific deployment.
- [OpenAI — gpt-oss-120b & gpt-oss-20b Model Card](https://openai.com/index/gpt-oss-model-card/) — Documents the gpt-oss model family, release conditions, evaluations, safety analysis, capabilities, and limitations alongside downloadable model weights. **Use:** First-party case study. **Limitation:** The model card is specific to two models and their release-time evaluations; it is not an enterprise acceptance decision or evidence for unrelated open-weight models.

---

This library synthesizes cited research, standards, official documentation, and clearly attributed practitioner perspectives. Maturity describes the state of a concept—not vendor endorsement, production readiness, or permission to deploy. Benchmarks and demonstrations do not replace use-case evaluation, governed controls, independent verification, or named human release authority.