Magic claims new recipe matches DeepSeek V4 Pro pretraining with 50x less compute (~$0.5M)
daniel_mac8 · x · 2026-09-09
- Magic (@magicailabs) pushed back on the idea that frontier pretraining is a big-lab-only game: without 100k chips, algorithmic efficiency is the only path.
- It claims its new recipe matches DeepSeek V4 Pro's pretrain using 50x less compute — roughly half the FLOPs used for GPT-3, or $0.5M on GB200.
- Caveat: company self-reported claim, no third-party verification yet.
Related event: Magic Claims 50x Pretraining Efficiency, Matching DeepSeek for ~$500K(6 posts)→
More from Infra
- Open-source Aurora: Go LLM gateway fork claims 55x LiteLLM speed and fixes OpenCode session lockout — entitybtw · 2026-09-09
- Navier-Stokes proof burned 300B output tokens, $20-30M at consumer API prices — soumitrashukla9 · 2026-09-09
- Nvidia and AMD fight over credit guarantees to bankroll customers' data centers — rohanpaul_ai · 2026-09-09
- DeepSeek 4 Flash runs all day on Spark: zero crashes, 98% cache hit, 40 tok/s — jasonkneen · 2026-09-09
- New report: custom ASIC market shifts from design wins to program responsibility — BenBajarin · 2026-09-09
- Uno speeds up Qwen3-8B 2.5x by using diffusion for parallel token drafting — rohanpaul_ai · 2026-09-09