Dossier
Open-model release record
A bounded record of what selected US and Chinese model developers released, licensed, and claimed about open-weight strategy. It separates inspectable artifacts from founder narratives and future promises.
Last reviewed 22 JUL 2026
What is established, developing, or disputed
Meta released Llama 2 weights and starting code under a custom community license for research and most commercial use.
ImplementedMeta expanded its open-model ecosystem pitch with Llama 3.1, including partners, safety tools, derivative-model permissions, and a proposed stack interface.
ImplementedLiang Wenfeng described open releases as part of DeepSeek's talent, research, and ecosystem strategy in a 2024 founder interview.
ReportedMoonshot AI says it will release Kimi K3 artifacts on July 27, 2026.
Officially announcedOpen-weight availability alone establishes lower cost per successful task or broader real-world adoption.
Contested
Principal entities
- People
- Liang Wenfeng · Mark Zuckerberg
- Organizations
- DeepSeek · Meta · Microsoft · Moonshot AI
- Places
- China · United States
Primary documents
Unresolved questions
- Do Kimi K3's promised artifacts and license appear on July 27 in the form described?
- Which license restrictions materially constrain commercial deployment or derivative models across these releases?
- How do serving cost, hardware requirements, and cost per completed task compare under reproducible workloads?
- Will weights, code, and documentation remain available and maintained over time?
Latest dispatches
Full chronology →Article · Meta AI · en
Llama 2 established the American open-weight precedentArticle · Meta AI · en
Llama 3.1 made Meta's open-model strategy explicitArticle · ChinaTalk · en and zh
Liang Wenfeng described DeepSeek's playbook before the shock