🧠 Microsoft AI finished pre-training MAI-Base-2 • MoE with 2.8T total / 80B active parameters • Natively multimodal (encoder-free vision) • Microsoft says it beats every open-weight model it tested and is 4–12x more FLOPs-efficient in pre-training than DeepSeek-V4-Pro-Base • Not released yet: it's in post-training and will power their next models
New model detected ·
mai-base-2
by Microsoft found in Microsoft AI newsroom
New model
Official source
- Status
- Released
- Maker
- Microsoft
- Found in
- Microsoft AI newsroom microsoft.ai
- Detected
- Posted to X
- 5 min after detection
- Model ID
mai-base-2
Evidence
Microsoft added 2 modelsMicrosoft AI newsroom: 2 added model entries
+mai-base-1
analysisMicrosoft AI newsroom reported a model catalog change on 2026-10-15 adding four models: mai-base-1, mai-base-2, mai-base-3, and mai-thinking-1. No models were removed. The source is the Microsoft AI news post "Frontier pre-training at Microsoft AI" (microsoft.ai/news/frontier-pre-training-at-microsoft-ai). The record lists release stage as "release" with no official preview or preview access details. No benchmark rows or scores were provided for these models.