Microsoft Pre-Trains MAI Base 2: 80B Active/2.8T Total MoE, 4-12x More FLOPs-Efficient Than DeepSeek-V4-Pro-Base
liuziwei7 · x · 2026-11-16
- Microsoft AI announced it completed pre-training MAI Base 2, a new state-of-the-art MoE foundation model with 80B active and 2.8T total parameters.
- It outperforms every open-weight model tested on held-out evaluations and is 4-12x more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base; the savings went into training a larger model.
- Follows MAI Thinking 1 (June), a general-purpose model post-trained on MAI Base 1 with strong reasoning performance.
- Researcher chillchio says the approach "works like a charm": slightly more training FLOPs than a vision-encoder setup, but worth removing encoder pre-training costs.
More from Models
- GPT-5.6 Sol helps uncover universal XSS in uBlock Origin allowing JS on any site — jedisct1 · 2026-11-16
- Model runs an ML experiment unprompted, complete with optimizer scaling analysis — teortaxesTex · 2026-11-16
- A curated list of 7 free uncensored local models, smallest just 0.67GB — nevrekaraishwa2 · 2026-11-16
- Rumor Speculation: Anthropic May Be Sitting on Ready Releases as OpenAI Lags Opus 5.5 — haider1 · 2026-11-16
- Coder's 'My life now' post on burning through weekly AI usage limits — Spare-Ad-1024 · 2026-11-16
- Screenshot appears to show unreleased "Opus 5.5 Medium" model in Claude Code UI — itsOmSarraf_ · 2026-11-16