Magic details >10x compute-efficient pretraining, eyes trillion-parameter models
Dr_Singularity · x · 2026-09-09
Magic published a research update on compute-efficient pretraining:
- Its recipe is >10x more compute-efficient than leading open-weight base models, matching DeepSeek V4 Pro Base with 50x fewer FLOPs (half of GPT-3's pretraining compute, $0.5M on GB200)
- Scaling 10x further ($4M) meaningfully beat all public open base models on perplexity; per its fitted scaling laws, DeepSeek's recipe would need >$100M for equivalent capability
- Method: bits-per-byte loss to normalize tokenizers, evaluated on heldout code/papers/math, covering open base models from DeepSeek, Moonshot, NVIDIA; notes issues measuring logprobs in vLLM/SGLang
- Strategy: pretraining + agentic RL + long context suffice for superhuman coding agents and automated AI R&D; scaling toward trillion parameters
Related event: Magic Claims to Match DeepSeek V4 Pro Pretraining with 1/50 the Compute(11 posts)→
More from Infra
- Anthropic's ~11 GW contracted compute estimated at ~42% NVIDIA GPUs, ~4.5 GW Google TPU — ns123abc · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- RTX 5080 price jumps from $1658 to $2318 in ten days on Amazon — Frizzy-MacDrizzle · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Neon rebuilds its docs search entirely on its own stack: Lakebase Search, Functions, AI Gateway — matei_zaharia · 2026-10-27
- PyTorch details 2.12x faster on-device streaming ASR on Arm SME2 via ExecuTorch — PyTorch · 2026-10-27