AI’s next race is about more than bigger models

MiniMax says its refreshed model brings major improvements, though the announcement details here don’t specify how those gains were measured. For builders, the useful question is whether the changes translate into better results on their own tasks—not just a stronger headline.

The model is reported to be available to Pro and Ultra subscribers in Google Search AI Mode, with daily limits and tiered pricing. That makes access and usage caps part of the product decision: teams will want to test whether the workflow justifies the subscription.

The new Daedalus system is rated at 85.69 petaflops, opening more room for advanced-model development in Greece. Local compute capacity can help research communities build expertise and pursue projects closer to home.

DeepSeek says the updated model completes tasks faster. If that holds up in real workloads, faster responses could make agent-style tools feel more usable—but teams should weigh speed against quality on their own evaluations.

The company’s new MCP server, built with Amazon Bedrock, gives client agents direct access to its cross-border payments datasets. It’s a useful example of how domain-specific data could make AI workflows more actionable, while putting access controls and data governance in focus.

InsurStaq AI’s new app is designed to inspect repositories locally rather than send source code to external servers, with a listed starting price of $19 per month. For developers working with sensitive code, the local-first approach may be as important as the findings themselves.

New Mexico State University says its system will support AI and machine learning research alongside work on water, wildfire, and energy. Shared university infrastructure can give researchers more room to explore problems that may not fit neatly into commercial AI priorities.

A 36Kr commentary argues that Apple has yet to produce an AI breakthrough with the broad impact of the iPhone or ChatGPT. For users, the bigger test may be whether AI becomes genuinely useful in everyday products, rather than arriving as a standalone spectacle.

Timestral ranges from 5 million to 209 million parameters and is presented as capable of zero-shot forecasting on a CPU. Smaller, specialized models could make forecasting more accessible where teams have limited hardware or don’t need a general-purpose assistant.

Sergey Brin has reportedly told key AI staff to go all in on Gemini and move faster. If that reflects a sharper internal focus, users may see more rapid product changes; developers will want to track how those changes affect APIs and workflows they rely on.