
Microsoft Claims 4–12× Efficiency Edge With MAI-Base-2
The 2.8-trillion-parameter model strengthens Microsoft’s in-house AI ambitions, but its benchmark claims remain internal and its weights are not publicly available.
Explore AI news, practical guides, tutorials, reviews, and insights from Zeniteq.

The 2.8-trillion-parameter model strengthens Microsoft’s in-house AI ambitions, but its benchmark claims remain internal and its weights are not publicly available.

The SDK supports Windows, macOS and Linux, but Windows session isolation ships ahead of Entra attribution and enterprise management integrations.

Microsoft’s Qwen-based scoring model offers inexpensive agent control in Foundry, while its headline speed and accuracy results remain vendor-reported.

Eligible Defender Suite and Microsoft 365 E5/E7 customers can try integrated security workflows, but workspace requirements and ingestion costs matter.

Maia 200 already serves Copilot workloads, while Microsoft plans multi-gigawatt deployments to make its accelerator family the main inference platform.
From Microsoft's internal pilot cancellation to Uber's blown AI budget, token-based pricing is forcing a reckoning across the industry.

The new white paper offers architecture guidance, while preview-stage Foundry Local documents how connected and disconnected on-premises inference can work.

Reported spending and usage reductions highlight platform owners’ incentives to favor their own tools, without establishing a break with Anthropic.

A Windows filename mismatch let hijacked Copilot agents alter protected settings in Visual Studio and SSMS, according to the researcher who disclosed…

Microsoft is combining collaboration, app building and persistent agents in Copilot, but each experience has a different access timeline.

The redesign brings Office work, natural-language app building and persistent agents together, but availability and billing differ sharply between them.

The $5,999 Dev Box offers a desktop alternative, but local-model capacity and forthcoming Windows agent controls need separate scrutiny.

Everyday Chat and Office work stays under a user license, but advanced agents require Copilot Credits and an admin billing policy.

Microsoft claims faster responses and fewer tokens than its June model, while GitHub lists usage rates developers can evaluate against completed tasks.

Azure developers gain Microsoft-managed deployment, enterprise authentication, and Azure billing for Grok 4.7, separate from the model’s September launch.

Apps and Managed Runtime enter public preview, while Foundry IQ and the Review panel offer generally available grounding and readiness checks.

Microsoft separates generally available CRM skills from public-preview sales and Teams features, while restricting its more autonomous Service Agent to selected customers.

Vegalabs says $21,168 in unused credits did not cover its Claude invoice; Microsoft explicitly excludes Anthropic models from startup-credit coverage.

Microsoft's experimental update lets Windows developers run GGUF models locally while keeping ONNX workflows and reusing familiar OpenAI client code.

Microsoft’s president wants enforceable safety duties, independent evaluations and containment controls, challenging Washington’s preference for voluntary commitments from AI companies.
Anthropic released a Claude add-on for Microsoft PowerPoint, but it wasn’t great. Here’s a better alternative.

Microsoft’s CEO proposes independent permissions, readable action logs and a human-operated emergency brake for advanced AI systems working with sensitive enterprise data.