Hugging Face has announced the open-sourcing of the Qwen3-4B model, which was distilled from 2.3 million reasoning traces of Claude Fable 5. The model achieved 100% self-consistency at 512 samples, with an output entropy of 0.00 bits and zero hallucination variance. This development indicates that the student model is not constrained by the teacher model, and the research has led to the convergence on a universal truth.
Hugging Face Open-Sources Qwen3-4B Model with Zero Hallucination Variance
More Articles From This Day
Macron and Modi Lead Global AI Infrastructure Race with Strategic Tech CEO Engagement
French President Emmanuel Macron and Indian Prime Minister Narendra Modi are actively courting technology CEOs to secure significant investments in AI infrastructure. Macron recently negotiated a €75 billion data center deal with SoftBank, facilitated through personal text exchanges with founder Masayoshi Son. Modi, following a similar approach, secured a $48 billion commitment from Amazon's CEO Andy Jassy for AI and cloud infrastructure in India. These engagements highlight the importance of direct relationships between heads of state and tech leaders in driving substantial investments that will shape the future of AI systems. The recent Choose France Summit and the India AI Impact Summit have showcased the growing stakes in AI infrastructure development, with billions pledged for projects in both countries.
