MAINLAND DISPATCH

Crosscurrents · Compare coverage

How three model releases define openness

Meta's Llama releases and Liang Wenfeng's pre-R1 interview use similar ecosystem language from different institutional positions. This comparison separates the artifacts they released from the strategic claims they made about them.

Mainland source

ChinaTalk / 36Kr · 27 NOV 2024 · EN AND ZH

Deepseek: The Quiet Giant Leading China’s AI Race

Moats created by closed source are temporary.

Framing

An annotated translation of a founder interview. Liang frames open release as a way to attract talent, accelerate an ecosystem, and make a temporary technical advantage more useful than secrecy would.

Open source ↗

Shared facts

  • All three records treat access to model weights as an ecosystem strategy, not only a research disclosure.
  • Each source is first-party or founder-originated and therefore documents a position more reliably than it proves outcomes.
  • None of the records makes the associated model unrestricted public-domain software.
  • Licenses, release artifacts, infrastructure costs, and derivative-use rights must be compared separately.

Differing emphasis

  • Llama 2 emphasizes availability for research and most commercial use; Llama 3.1 expands the argument to partners, tools, safety, and distribution.
  • Liang emphasizes the temporary nature of technical advantage and the recruiting value of an open ecosystem.
  • Meta labels its strategy open source, while the underlying custom licenses retain conditions that make open weight the more precise cross-release term.
  • The founder interview describes intent; the Meta release posts provide more directly inspectable product and license artifacts.

Editorial notes

This is a comparison of release posture, not a nationality score or a capability ranking. The sources establish what the organizations published and said. They do not, by themselves, establish adoption, cost per completed task, safety, or the durability of access.

Dispatches in this comparison