DigiNews

Tech Watch by Johan Denoyer

← Back to articles

>10x More Efficient Pretraining

Quality: 8/10 Relevance: 9/10

Summary

Magic AI, Inc. reports a >10x improvement in compute efficiency for pretraining trillion-parameter models, achieving ~50x fewer FLOPs than DeepSeek V4 Pro Base and strong perplexity gains. The update covers scaling laws, evaluation across multiple domains, and plans for long-context RL and alignment research.

🚀 Service construit par Johan Denoyer