rbtfl

Open-weight licensing gets a standard, and the G7 a spectrum

Linux Foundation ships OpenMDW-1.1 (Nvidia adopts for four model families); days later the G7 replaces the binary open/closed split with a spectrum

AI· transition Who Decides·Whose Money ·9 takes · ·rbtfl upd Jun 24, 2026
post

The split

The same story, as told by newsrooms in different countries. Their words, attributed and linked.

United Kingdom

Diginomica

“OpenMDW-1.1 conveys rights across model, data and weights as one artifact, the contract clarity an Apache license bolted onto AI weights can't give enterprises.”

enterprise-software analysisread the original ↗

United States

Open Source Initiative

“Open weights are not open source: downloadable parameters without training code or data information fall short of the Open Source AI Definition.”

open-source steward / definitionalread the original ↗

post

Summary

Open-model licensing, long a legal minefield of software licenses bolted onto weights, is getting standardised. On 28 May 2026 the Linux Foundation released OpenMDW-1.1 (Open Model, Data and Weights), built with Amazon, Meta, IBM and Microsoft, which defines the rights conveyed for whatever a publisher chooses to share, rather than gatekeeping what counts as "open." Nvidia adopted it across Cosmos, Isaac GR00T, Ising and Nemotron. Days later, on 2 June, the G7 Digital and Technology Ministers published a framework replacing the binary open/closed split with a spectrum, effectively blessing "open weights" as the term of art. The Open Source Initiative dissents: downloadable weights without training code and data information are not open source under OSAID.

By the numbers

  • 28 May 2026, OpenMDW-1.1 released (Linux Foundation).
  • 4, Nvidia model families adopting it (Cosmos, Isaac GR00T, Ising, Nemotron).
  • 2 June 2026, G7 framework swapping the open/closed binary for a spectrum.
  • 2024, year OSI's Open Source AI Definition (OSAID) set the contested bar.

Why it matters

License ambiguity has kept regulated enterprises off open-weight models even as DeepSeek, Llama and Qwen reshape the frontier. A clean rights-grant plus a G7 vocabulary lowers the legal cost of building on open weights, while the OSI's objection signals the fight over what "open" means is being settled by industry, not stewards.

What to watch

  • Whether other labs (Meta, Mistral, Alibaba) standardise on OpenMDW.
  • How national regimes operationalise the G7 "spectrum" in AI rules.
  • OSI vs industry over the "open source" vs "open weights" label.

The briefing, by email