Terminology and licensing guide

Open Source vs Open Weight AI Models

“Open source model” is common search language, but AI model openness has multiple layers. This guide separates open source AI, open-weight releases, weights available under restricted terms, and closed API-only models so teams can evaluate licenses and deployment risk accurately.

Open Source AI

The model and accompanying information grant the freedoms described by the current Open Source AI Definition.

Open Weight

Weights are published, but license terms, training data transparency, code, redistribution, or usage freedoms may be narrower.

Restricted Weights

Weights are available but gated or governed by custom terms, acceptable-use policies, or redistribution limits.

Closed/API-only

The model is accessed through a provider API and the underlying weights are not published for independent use.

Open-source coding software

This describes the agent, editor extension, CLI, runtime, or server code and the license governing that software. For example, Kilo's VS Code extension, JetBrains plugin, and CLI are MIT licensed.

See Kilo's component and license matrix

Open-weight model artifacts

This describes access to a model's learned weights under stated terms. It does not, by itself, establish that the complete AI system meets an open-source definition or that every commercial, modification, and redistribution right is unrestricted.

View evidence-reviewed open-weight models

What To Review

Code license

Is inference, training, tokenizer, or serving code included and reusable?

Weight license

Can you use, modify, fine-tune, redistribute, or host the weights commercially?

Data transparency

Does the release provide enough information to understand training data and limitations?

Reproducibility

Can another party inspect or reproduce meaningful parts of the system?

Auditability

Can security, privacy, and compliance teams evaluate model behavior and supply chain?

Procurement fit

Do license obligations, usage restrictions, indemnity gaps, or export controls affect your deployment?

Registry Examples

These examples come directly from the shared evidence registry. They illustrate terminology and license review, not a separate ranking. Status was last verified 2026-07-30.

Artifact	Classification	License	Weight status	Caveat
GLM 5.2	open-weight	MIT	available	Our current top editorial pick because its official release combines strong coding-agent results, a 1M-token context window, published weights, and permissive MIT terms.
Qwen3 Coder 480B-A35B Instruct	open-weight	Apache-2.0	available	A coding-specific checkpoint with published weights, Apache-2.0 terms, native 262K context, and first-party agentic coding guidance.
Kimi K3	open-weight	Modified MIT (Kimi K3 License)	available	A long-horizon agentic-coding model with published weights, Modified-MIT license terms, a native 1M-token context window, and vendor-reported terminal-agent coding results.
DeepSeek V4	open-weight	Model-card license	gated	Retained in the evidence registry, but excluded from the current ranking until its gated artifact, license, and benchmark evidence can be reviewed again.
Devstral Small 2505	open-weight	Apache-2.0	available	The official release documents a 24B Apache-2.0 checkpoint that can run on a single RTX 4090 or a Mac with 32 GB memory.
Laguna S 2.1	open-weight	OpenMDW-1.1	available	A compact MoE coding model with published weights, the permissive OpenMDW-1.1 license, a native 1M-token context window, and vendor-reported results competitive with larger models.
Llama 3 8B Instruct	weights-available	Meta Llama 3 Community License	gated	Included on local and terminology pages as weights-available under custom terms, not as an open-source coding-model winner.

Primary References

Open Source AI Definition G7 code of conduct for advanced AI systems

FAQ

Is an open-weight model automatically open source?

No. Open-weight means model weights are available under stated terms. Open source AI requires broader freedoms and transparency; code, data information, license terms, and redistribution rights still matter.

Do open source AI definitions require publishing the raw training dataset?

No. Follow the current Open Source AI Definition rather than assuming raw dataset publication is always required. The definition focuses on the information and freedoms needed to understand, use, modify, and share the system.

Is open-source software the same as an open-weight model?

No. Software licensing covers code such as an AI coding agent or inference runtime. Model licensing covers artifacts such as weights, model code, and associated terms. An MIT-licensed coding agent can use a closed API model, and an open-weight model can be served through proprietary software.

Why does procurement care about the difference?

Commercial use, redistribution, fine-tuning, acceptable-use restrictions, data transparency, and auditability can differ even when weights are downloadable.