Event
DeepSeek-R1 untersucht mehrschrittige Aufgaben
A research paper published in January 2025 on DeepSeek-R1 studies, among other things, reinforcement learning for language models tackling multi-step tasks. The authors also report releasing models and smaller distilled variants. Source
At the time of access, DeepSeek’s repository described its weights as MIT-licensed while also noting the original licenses of the Qwen and Llama base models used for distilled variants. The Llama 8B variant’s metadata says MIT, while its text refers to the base model’s Llama 3.1 license. The relationship between these statements remains open here; releasing weights does not imply releasing all training data. License statements
Sources
- Forschungsarbeit: Ansatz und Angaben zur Freigabe.
- Lizenzangaben zu Varianten und Basismodellen: aktueller Dokumentationsstand und offene Auslegung.
Offen
Historische Lizenzstände, das Verhältnis der Llama-Lizenz zur MIT-Angabe und berichtete Kennzahlen weiter prüfen. / Further investigate historical license versions, the relationship between the Llama license and MIT label, and reported metrics.