Hi Dotikers!

Meta is back in the open model game, and this time it actually means business. The company just released Muse Spark Open, an 816 billion parameter model (42B active) that it calls the most capable US open-weight model to date. The weights are available everywhere that matters: Hugging Face, SageMaker, Google Cloud, plus a long list of API providers from Groq to Together AI.

Let's be honest, the "open" label deserves some side-eye. The Muse Model License requires any company making over $20 million in revenue to sign a separate agreement with Meta before offering the model as a service, and everyone must display "Built with Muse Spark Open" in their product. That's not open source, that's open-ish with a marketing clause attached. Meta is essentially copying the playbook of Moonshot's Kimi K3 license, which is ironic given how loudly the industry complained about Chinese licensing terms just a year ago.

On the performance side, the picture is genuinely impressive but incomplete. Spark Open crushes its American rivals, Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra, often by 20 points or more. Against the Chinese heavyweights, it leads on agentic and tool-use benchmarks but still trails Kimi K3 and GLM 5.3 on coding tasks like Terminal-Bench. Meta acknowledged this gap back in April with the closed version, and it clearly hasn't closed it since. For a model explicitly built for software engineering, that's the one benchmark where excuses don't fly.

Still, this release matters. It puts real competitive pressure back on the open-weight ecosystem from an American lab, something developers have been waiting for. And if you want to run it locally, all you need is two maxed-out Mac Studios, which cost roughly the same as a small car but depreciate even faster.

More on what this means for the open ecosystem below.

Alex.

Agents now issue refunds, update accounts, and close tickets end to end. Running Agents in Customer Work: four Tuesdays on agentic AI in customer ops. Register for All Four