Microsoft's MAI-Base-2 scaling ladder pegs the L42 run at ~1.8×10^23 FLOPs
teortaxesTex · x · 2026-10-23
The attached figure from Microsoft's MAI-Base-2 report shows a scaling ladder built from intermediate checkpoints (L18–L42) that accurately extrapolates the final model's performance, letting the team isolate training-recipe efficiency from raw compute — using DeepSeek-V4-Pro-Base as the external baseline.
teortaxesTex comments that Microsoft seemingly concluded the L42 run needed 1.8×10^23 FLOPs with a very conservative architecture, and speculates DeepSeek's compute budget might be overstretched if it runs such ladder experiments for every candidate. He adds that V4.1F is already over an order of magnitude larger than V4.1, so smaller-size extrapolation curves would help.
More from Models
- Meta researcher: horrifying infra, not talent, is why OSS can't catch closed frontier — A_K_Nain · 2026-10-23
- TypeSafe's Jev: a "Decision Model" LLM That Returns Floats at $0.042/M Input Tokens — ariG23498 · 2026-10-23
- Local model agentic coding benchmark: 27B matches 125B, tool-calling is the lever — Joralk_ · 2026-10-23
- NVIDIA open-sources PersonaPlex-7B, a full-duplex voice AI model that listens and talks simultaneously — solyarisoftware · 2026-10-23
- NVIDIA's KumoRFM-2: a foundation model that predicts churn, fraud and demand straight from relational databases — solyarisoftware · 2026-10-23
- Gemini Analyzes Whether a Simple 'Hi' Is a Jailbreak Attempt in Its Thinking Steps — OldIntroduction2909 · 2026-10-23