APIUp to 25% cheaper than official pricesTry the API →
Hermes
APIModelsCases
Back to News
newsTraining InfraLlms

Magic Matches DeepSeek V4 Pro Base Using 50x Less Compute

Magic claims a pretraining recipe that matches DeepSeek V4 Pro Base with roughly 50x fewer FLOPs, hitting frontier quality for around $0.5M on GB200.

Magic Matches DeepSeek V4 Pro Base Using 50x Less Compute
Source
Magic
Published
Author
AlphaSignal Newsroom
Read
1 min read

Magic claims a pretraining recipe that matches DeepSeek V4 Pro Base with roughly 50x fewer FLOPs, hitting frontier quality for around $0.5M on GB200.

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Read original report

Next reads

Sewon Min's Preprint Shows Smaller Batches Beat Overfitting on Repeated Training Data
AlphaSignal · paper

Sewon Min's Preprint Shows Smaller Batches Beat Overfitting on Repeated Training Data

MSAR Cuts Transformer Training Compute by 37% With Smarter Residual Connections
AlphaSignal · paper

MSAR Cuts Transformer Training Compute by 37% With Smarter Residual Connections

Zyphra Cuts MoE Training Communication Costs by 2.63x Using Router Patterns
AlphaSignal · paper

Zyphra Cuts MoE Training Communication Costs by 2.63x Using Router Patterns