Meta's open frontier model still runs on NVIDIA's terms

Meta's new open model is free to download and takes about 1.5 terabytes just to load at full precision. According to the Hugging Face write-up of Muse Spark Open, it is a mixture-of-experts model with roughly 816B parameters. A router picks 8 of 256 experts for each token, so only about 42B parameters do the work. You get roughly the speed of a 42B model, but every expert stays in memory. The post lists 4 nodes of 8 H200 GPUs for the BF16 checkpoint. A 4-bit NVFP4 version fits on a single B200 node, with what the authors call negligible accuracy loss. A full 1M-token context adds about 266 GB of cache on top.

This is good news for anyone who cares where their models run. With Muse Spark Open and Mistral Large 4 (1T parameters, 49B active, by Mistral's count), a European company can host a frontier-class coding agent in its own data center and fine-tune it there. No API provider can reprice it or retire it. But the affordable deployment path is a format built specifically for NVIDIA Blackwell. The weights are open, while the hardware you need to run them cheaply comes from one company. Sovereignty planning that stops at the model license is missing half the stack.

For teams building agents, one detail matters more than the headline benchmarks: reasoning cannot be switched off. The lowest setting still spends tokens thinking on every step of a long tool-calling loop. Meta calls this the most capable US open-weights model, and that is a claim until it is tested in harnesses other than Meta's. My open question is whether the 4-bit checkpoint holds up over long debugging sessions, where small accuracy losses have more steps in which to add up.