Open-weight tells you about access to parameters. Open-source asks a larger question.
An open-weight model makes trained parameters available under some license. That can be enough to run inference, inspect layers, fine-tune a checkpoint, or deploy locally—if the license permits those actions. It does not, by itself, provide the data information, training and evaluation code, documentation, or legal freedoms needed to understand and modify the complete AI system.
The Open Source Initiative’s Open Source AI Definition evaluates whether people can use, study, modify, and share an AI system, including access to the preferred form for making modifications. That is broader than “the weights can be downloaded.” The definition gives open-source AI a testable meaning instead of treating “open” as a mood.
| Question | Open-weight release | Open-source AI claim |
|---|---|---|
| Are trained parameters available? | Usually yes, under stated terms. | Parameters are one required component. |
| Is meaningful training-data information available? | Not necessarily. | Required in a form that supports studying and modifying the system. |
| Are training and inference code available? | May include only example inference code. | Required code must support exercising the relevant freedoms. |
| Can users modify and share? | Depends entirely on the model license. | Qualifying freedoms and terms are central to the claim. |
| Can the model be reproduced exactly? | Usually not from weights alone. | Openness improves modifiability; exact reproduction can still be constrained by compute or unavailable raw data. |
Inspect the release as a stack of artifacts.
Weights and configuration
Checkpoint format, architecture configuration, tokenizer, immutable revision, checksums, precision, and derived variants.
Training and inference
Training loop, preprocessing, evaluation, serving examples, dependencies, build instructions, and remote-code requirements.
Data information
Sources, curation, filtering, licensing, synthetic generation, deduplication, exclusions, governance, and known gaps.
Documentation and results
Model card, system card, intended uses, limitations, benchmarks, prompts, settings, hardware, risk work, and change history.
A release can be valuable even when it is only open-weight. Precision matters because developers, researchers, and buyers make different decisions when they expect the full development form instead of a runnable checkpoint.
The license decides what you may do—not the repository badge.
Before research or deployment, answer these questions from the exact license and release terms:
- May the organization use the model commercially or offer it as a hosted service?
- May weights, quantizations, adapters, merged checkpoints, and other derivatives be redistributed?
- Do attribution, notice, naming, branding, or disclosure obligations apply?
- Are there user-count, revenue, field-of-use, geography, or prohibited-use restrictions?
- Do downstream recipients receive the same rights, and which files must travel with a derivative?
- Are the tokenizer, code, and bundled datasets under separate licenses?
This is a technical checklist, not legal advice. For a production deployment, preserve notices and have qualified counsel review the exact upstream revision and every derivative or quantization in use.
PRIMARY SOURCES
- 01Open Source AI Definition 1.0
Open Source Initiative — Formal definition and required components for open-source AI.
- 02Model Openness Framework
LF AI & Data Foundation — A framework for grading completeness and openness of model releases.
Openness, modifiability, and reproducibility overlap—but they are not identical.
A system can grant meaningful freedoms without making an exact training run affordable to repeat. Conversely, a paper can describe a reproducible method while the resulting checkpoint carries restrictive terms. Track the claims separately: access to artifacts, legal permissions, technical modifiability, and the practical ability to recreate results.
For model evaluation, reproducibility also requires the served version, precision, quantization, system prompt, sampling settings, tool or retrieval scaffolding, and scoring method. Two endpoints using the “same” base model can behave differently because the deployed systems are different.
BrokenGPT uses the narrower label when full openness is not established.
BrokenGPT is an India-built chat and API product, not the upstream foundation model. In Phase 1 it can serve configured third-party or open-weight models. Calling that product an “open-source model” would collapse the model, its release license, and the hosted service into one inaccurate phrase.
The editorial standard is therefore to say open-weight when weights are the verified open artifact, and to reserve open-source AI for releases that can be evaluated against a formal definition. The model disclosure remains the place to document the active configuration, provenance, obligations, and limitations.