
Alibaba Open-Sources 30B Multimodal Model: 3 Billion Activated Parameters, Taking On GPT-5-Mini Head-On
With 3 billion activated parameters, it delivers near-flagship multimodal capability.
Category
327 articles

With 3 billion activated parameters, it delivers near-flagship multimodal capability.

The cost was always quoted per generation. The real cost is priced per asset that survives review.

Capability is largely there. The economics are not. Forty years to ten per cent is not a forecast that justifies panic.

The licence is the marketing. A 744B model anyone can download resets the buy-versus-build calculation.

What's going global isn't the series, but the tools for making series.

A small release with a large implication. Agent behavior stops being a setting you configure and becomes something you write.

A price cut becomes a product change when it moves generation from a decision into a reflex.

The organizations seeing results picked a process, measured it, and rebuilt it. The ones stuck in pilots bought access and waited.

A control that fires in seconds is a postmortem tool. A control that fires in milliseconds is a seatbelt.

The competitive focus of open-source image generation is shifting from looking realistic to running feasibly and editing precisely.

There is no single best voice model yet. The right pick depends on whether your bottleneck is hearing in noise, finishing a task, or answering fast enough.

When a model changes one thing and leaves the rest alone, it stops being a novelty and becomes a step you can chain. The money is in the workflow, not the picture.