AI digest: Open models fight back
Google releases Gemma 4 under Apache 2.0, Microsoft launches three new models, and the battle for local AI heats up.
The open source movement is making serious moves this week. Google’s gone full Apache 2.0 for the first time, whilst Microsoft counters with three new foundational models.
Google’s Gemma 4 goes properly open
Google released Gemma 4 under Apache 2.0 licensing, marking the first time they’ve used fully open licensing instead of their restrictive custom terms. The four models run everything from phones to workstations, which matters because it means you can actually build commercial products without worrying about Google’s lawyers. This is a proper shot across the bow at OpenAI’s closed approach.
Microsoft fires back with MAI models
Not to be outdone, Microsoft launched three new foundational models that handle voice transcription, audio generation, and image creation. The transcription model runs 2.5x faster than before at £0.27 per audio hour, which is proper competitive pricing. Microsoft’s clearly betting on multimodal being the differentiator, though we’ll see if enterprises care more about speed or capabilities.
Trinity Large Thinking goes open source
Arcee AI dropped Trinity Large Thinking as an Apache 2.0 reasoning model specifically built for long-horizon agents and tool use. This matters because it’s the first open reasoning model that’s actually designed for agent workflows rather than just chat. The timing’s perfect as everyone’s moving beyond simple Q&A towards autonomous systems.
Local AI gets serious hardware backing
NVIDIA and Google are pushing hard on local agentic AI to dodge the “token tax” of cloud APIs. Running Gemma 4 locally on RTX hardware means you can build always-on assistants without bleeding cash on API calls. This could actually change the economics of AI deployment for smaller companies.