AI Pulse · Frontier Labs Watch

Meta just released Muse Glimmer, a 30-billion-parameter AI model you can run on a single consumer GPU. It’s open-weight, licensed under Apache 2.0, and built specifically for agentic workflows that need to stay running locally. According to the original report, it decodes 3.1 times faster than comparable models and works with vLLM from day one, which matters if you’re planning to deploy anything that needs to respond quickly without sending data to a third party.

The timing wasn’t subtle. The same day, Zuckerberg published an essay arguing that the US should remove barriers to open-source AI development. He named OpenAI, Anthropic, and Google as closed labs and pointed to Chinese competitors like Kimi K3, Qwen3.8-Max, and DeepSeek V4-Flash as the ones actually winning the open-weight race. The argument is that closed models create dependency, open models create optionality, and right now the most capable open-weight work is happening outside the US.

For anyone running a business, this matters in two ways. First, a 30B model that runs locally on affordable hardware changes the cost structure of automation. You’re not paying per token or worrying about rate limits. You can build always-on agents that handle repetitive tasks without sending proprietary data anywhere. Second, the political fight over open versus closed AI will shape what tools you can use and how much they cost. If open-weight models keep improving at this pace, the lock-in risk of relying entirely on closed APIs gets harder to justify.

This is exactly the kind of shift we track in the Omni Command Centre, where you can monitor which models are viable for specific workflows and what the trade-offs actually look like in production. The gap between open and closed is narrowing faster than most people expected.