MagicAILabs claims new recipe matches DeepSeek V4 Pro pretraining with 50x less compute, ~$0.5M
AccBalanced · x · 2026-09-09
MagicAILabs says frontier pretraining isn't a big-lab-only game: lacking 100k chips, they bet on algorithmic efficiency. Their new recipe reportedly matches DeepSeek V4 Pro's pretraining using 50x less compute — roughly half the FLOPs used for GPT-3, or $0.5M on GB200. Unverified by third parties, but if true it would drastically lower the barrier to frontier-scale pretraining.
Related event: Magic Claims 50x Pretraining Efficiency, Matching DeepSeek V4 Pro at ~$500K(7 posts)→
More from Infra
- Inside fal's H3 Max Director: streaming video generation with mid-stream prompt edits — noahsolomon · 2026-09-09
- SpaceX Buys Gas Turbine Business to Tackle Power Bottleneck on Road to 10GW — DMaguireARK · 2026-09-09
- vLLM's Hybrid HiSparse keeps decoding past HBM limits: 19-25 vs 5-6 concurrent 1M-context requests — vllm_project · 2026-09-09
- $22.5M of Compute in One Week: The New Price Tag for Math Breakthroughs — vykthur · 2026-09-09
- $1M sounds like a lot — it's just 10.5 minutes of OpenAI's compute spend — wordgrammer · 2026-09-09
- SageMaker Feature Store adds UpdateRecord for atomic feature-level writes, no more read-modify-write — AWS ML Blog · 2026-09-09