Microsoft claims higher training efficiency for MAI-Base-2

Published Thursday, October 15, 2026 · 11:42 PM EDT

Microsoft unveiled MAI-Base-2, claiming 4x-12x greater pretraining efficiency measured in floating-point operations than DeepSeek-V4-Pro-Base. The mixture-of-experts model has 2.8T total parameters and activates 80B, using a subset of its parameters for computation. The efficiency comparison is Microsoft's claim.

Sources

Techmeme

In this story

Microsoft DeepSeek

Published by Tech & Business, a media brand covering technology and business. This story was sourced from Techmeme and reviewed by the T&B editorial agent team.

Back to Newswire

Keep reading

Full wire

AI Science

Study finds cheaper claim verification carries higher unsupported-claim acceptance

Researchers report in a preprint that combining a cheaper claim verifier with a frontier language model scored 74.10 at USD 2.92 per thousand claims, against 73.83 at USD 8.93 for the frontier model alone. The system sends some cl...

AI Science

Researchers report lower-cost training for LLM information gathering

Researchers report in a preprint that training language models to select questions using information-gain rewards improved performance on Twenty Questions at substantially lower computational cost than the experimental-design base...

Science Security

Researchers report 66 faults in six software security testing tools

Researchers report that LLM4SAST, a framework for testing software security scanners, identified 66 unique faults across six tools. Developers confirmed or fixed 32 after receiving reports, according to the arXiv preprint. The fr...

Science

Researchers report lower quantum circuit resource needs with LLM-guided search

Researchers report in an arXiv preprint that a search framework using large language models discovered quantum circuits requiring fewer Toffoli gates and qubits than prior designs in seven of eight tasks. The models propose circui...

Capital AI

CarbonAI raises $12M for commercial property lending agents

CarbonAI has raised $12M in a round led by Cowboy Ventures for AI agents that handle commercial real estate loan underwriting, due diligence and monitoring after closing. Timber Grove Ventures, Bling Capital and Max Ventures also ...

AI Security

OpenRouter launches Armor beta to screen agent tool calls

OpenRouter says it has launched Armor in beta to check whether an AI agent's tool calls violate policy before execution. The service uses Typesafe AI's Jev model to screen those actions. OpenRouter claims better performance than ...