A review of open-weight model releases shows OpenMDW spreading across builders, modalities, and scientific domains, and provides evidence that a broader open AI ecosystem is taking shape.
When the Linux Foundation and PyTorch Foundation introduced OpenMDW in 2025, the premise was straightforward: machine learning models needed licensing designed for machine learning models. OpenMDW is a single, easy-to-apply permissive license enabling openness, consistency, and clarity across all components of an AI model distribution.
A modern AI model distribution can include architecture, weights and parameters, software, configuration, documentation, evaluation materials, training data, and other artifacts. Traditional software licenses were developed primarily around source code. OpenMDW provides a single licensing framework that can be applied across the “Model Materials” an organization chooses to release.
On May 28, 2026, that effort reached an important milestone with the release of OpenMDW-1.1 and NVIDIA’s announcement that it would adopt the license across future releases of its Cosmos, Isaac GR00T, Ising, and Nemotron open model families. Those families span agentic AI, robotics, simulation, and quantum computing.
That announcement raised a useful question:
How broadly is OpenMDW actually being used today?
We went looking.
OpenMDW adoption is broader than the headline numbers suggest
We reviewed model hubs, publisher repositories, model cards, license files, and release announcements to identify publicly available model weights actually released under OpenMDW-1.0 or OpenMDW-1.1.
The current census identifies 48 distinct upstream model and release groups across 12 publisher groups, spanning 36 AI and scientific domains.
OpenMDW-licensed weights now span:
- General-purpose and reasoning language models
- Coding and software engineering models
- Translation
- Reward, evaluation, and distillation models
- Text and multimodal embeddings
- Multimodal retrieval, reranking, and RAG
- Speech recognition and conversational audio
- Image and video generation
- Autonomous driving
- Robotics and world models
- Medical and surgical AI
- AI agent security
- Quantum computing
- Chemistry
- Weather and climate
- Geospatial and environmental modeling
NVIDIA represents a major portion of recent adoption, with 30 qualifying release groups in the current index spanning language and reasoning, world models, robotics, autonomous driving, retrieval, quantum computing, speech, safety, weather and Earth systems, and medical AI.
The adoption story extends considerably beyond NVIDIA, with several prominent neolabs adopting OpenMDW.
- Arcee AI moved its Trinity family to OpenMDW-1.1, including Trinity Nano, Mini, Large, Large-Thinking, and associated releases. Arcee specifically cited the value of having a model-oriented license covering the different materials contained in an AI model release. That work has since taken on additional significance. In August, the U.S. Department of Energy named Arcee the first industry partner supporting its new Genesis Open Models Initiative and announced Genesis-Science-1, the initiative’s first open-weight model for scientific research. The trillion-parameter-class model is being built with DOE and its national laboratories on Arcee’s next generation of Trinity models and is planned for open release.
- Poolside has released its Laguna 2.1 coding models under OpenMDW-1.1, bringing the license into another highly active area of open AI development: models designed for software engineering and coding agents.
- Earlier OpenMDW-1.0 adoption includes BAAI’s OpenSeek-Small family, where the repository itself contains the OpenMDW-1.0 license, and other releases across language, translation, multimodal, speech, and specialized scientific applications.
The emergence of initiatives such as DOE’s Genesis Open Models also shows the potential reach of this broader movement: open-weight models are becoming infrastructure for both commercial AI development and national-scale scientific research. The OpenMDW ecosystem is developing organically across organizations, geographies, model architectures, and use cases.
Counting OpenMDW license adoption turns out to be surprisingly difficult
Simply counting repositories produces misleading results. Model hubs contain community quantizations, mirrors, conversions, experimental checkpoints, and multiple encodings of the same underlying model. A single trained model might appear in several technical formats (e.g., BF16, FP8, GGUF) representing different precision levels or runtime targets for the same model.
The reverse problem also occurs. License metadata can disagree with the actual license contained in a repository. BAAI’s OpenSeek-Small-v1 provides a good example: repository metadata has at times identified another license while the repository’s actual LICENSE file contains OpenMDW-1.0.
Family-level announcements require similar care. NVIDIA’s May announcement covered future releases across four major families. Individual earlier models within those families do not automatically become OpenMDW models.
For this census, we therefore focused on upstream releases where the model weights themselves can be affirmatively tied to an OpenMDW license. Community mirrors and third-party quantizations do not become separate examples of adoption simply because they reproduce the same underlying model.
Why model-specific licensing matters
Open source software has benefited enormously from a relatively small number of well-understood licenses. Developers recognize Apache, MIT, BSD, GPL, and other common OSI approved license families and understand the basic framework under which software can be used and redistributed.
AI models introduced a different kind of artifact. A model distribution can combine rights and materials that do not map neatly onto a source-code repository. OpenMDW addresses that directly by defining Model Materials around machine learning models, including architecture and parameters, together with related artifacts the provider chooses to make available under the license.
OpenMDW-1.1 is designed to provide broad rights to use, modify, and redistribute those Model Materials and also explicitly addresses model-generated outputs. The practical benefit is consistency.
- A model developer gains a license designed around what is actually being distributed.
- A downstream developer gains greater clarity around the rights associated with those materials.
- An enterprise evaluating hundreds or thousands of AI components gains a more recognizable framework for compliance.
- An ecosystem gains common legal infrastructure that makes collaboration easier.
That type of shared infrastructure has played an important role throughout the history of open source.
The model layer is part of a much larger open AI stack
The emergence of OpenMDW also highlights something happening across the Linux Foundation ecosystem.
AI is rapidly developing into a stack of independently evolving technologies: models, training systems, inference engines, data infrastructure, agent frameworks, interoperability protocols, orchestration, cloud infrastructure, networking, observability, security, and governance.
Increasingly, important pieces of that stack are being developed in open communities under Linux Foundation governance. The PyTorch Foundation has evolved into a multi-project home that today includes PyTorch, vLLM, DeepSpeed, Ray, Helion, and Safetensors, spanning foundational model development, distributed computing, inference, optimization, and model distribution.
The Agentic AI Foundation (AAIF) is bringing together the standards and projects required for interoperable agents and critical components needed to deliver agentic applications reliably and securely. Its ecosystem includes MCP, AGENTS.md, goose, agentgateway, and Agent2Agent (A2A), the open standard for communication and collaboration between independent agents. A2A joined AAIF as a hosted project in August 2026, bringing another foundational interoperability protocol into the community. AAIF already describes its mission as building the open standard agentic AI stack, with participation from organizations across enterprise technology, cloud infrastructure, government, academia, and AI development.
The Cloud Native Computing Foundation (CNCF) provides much of the operational substrate on which AI workloads increasingly run. Kubernetes sits at the center, alongside projects and initiatives such as Kubeflow, KServe, Envoy AI Gateway, kagent, llm-d, and emerging AI infrastructure work across scheduling, model serving, networking, and agent infrastructure. The connection with the broader open AI ecosystem across LF is increasingly direct: for example, KubeRay, part of the Ray ecosystem under the PyTorch Foundation, provides the Kubernetes-native operator used to deploy and manage Ray workloads on Kubernetes.
And LF AI & Data hosts projects across the data and application layer, including Docling, Data Prep Kit, BeeAI, Delta Lake, OPEA, Open Model Initiative, ONNX, and many others supporting data preparation, model interoperability, enterprise AI, and applied AI workflows.
OpenMDW adds another important piece: common licensing infrastructure for the model layer itself.
Taken together, these communities increasingly span the lifecycle from model development through production operation:
Models + data → training → inference → agents → interoperability → deployment → observability → security
That convergence creates an opportunity larger than any individual project or foundation. It gives developers and organizations a neutral place to collaborate on the technologies that will define the open AI stack.
Building a living picture of the open model ecosystem
The census behind this post is a snapshot. The model ecosystem is moving too quickly for a static list to remain useful for long. We believe there is value in maintaining a living OpenMDW model index that can track qualifying upstream releases over time. At a minimum, that dataset can capture:
Publisher | Model | Model type | Parameters | OpenMDW version | Release date | Upstream or derivative | Additional terms | Source
Making that information publicly accessible and machine-readable would serve several communities at once.
Model developers could see how peers are applying OpenMDW. Enterprises could use it as an input into model discovery and compliance workflows. Researchers could study the evolution of open-model licensing and adoption. The wider community could help identify releases, correct metadata, and improve the quality of the dataset. Open ecosystems become stronger when their infrastructure is transparent and reusable. The same principle should apply to the information we use to understand those ecosystems.
Where this goes next
Open models are becoming an increasingly important part of the AI landscape. As that happens, the technical infrastructure around them will continue to mature and so will the legal, governance, and interoperability infrastructure required to build with them confidently.
OpenMDW is still early, but model developers are already applying it across very different kinds of AI, from language and coding models to autonomous systems, medical AI, agent security, multimodal retrieval, chemistry, weather and climate, and quantum computing. That breadth of adoption is an early indication that OpenMDW can serve as common licensing infrastructure across very different categories of AI.
That said, the growth of an open ecosystem for AI ultimately depends on more than access to weights. It depends on open data, open frameworks for building models, open infrastructure for operating them, open standards that allow systems to communicate, and neutral governance that lets competing organizations collaborate on shared technology.
Across models, frameworks, inference, data, agents, interoperability, and cloud infrastructure, a substantial portion of the open AI stack is now taking shape within Linux Foundation communities. OpenMDW adds common licensing infrastructure for the model layer, complementing the open frameworks, infrastructure, standards, and projects being developed across Linux Foundation communities and extending LF’s role in the open AI ecosystem.
Help us keep this OpenMDW living index current
- Released an open-weight model under OpenMDW? Let us know.
- Found a model we’ve missed or metadata that needs correcting? Open an issue
- Interested in adopting OpenMDW for your next model release? Learn more.