What happened
Meta published the weights for Muse Spark Open. It is natively multimodal, with 816 billion total parameters, 42 billion of them active at a time, and a context window of one million tokens.
Meta's pitch is that this is the most capable open-weight model the US has produced so far. Its AI chief, Alexandr Wang, said as much on X. The model is tuned for coding and agentic work: long tasks, serious software engineering, multi-step reasoning.
The headline detail is the hardware. Meta says the model runs on a desk using two Apple M5 Ultra Mac Studios and the MLX framework. Meta also says it is the first model in its class to do that.
Access is broad. API providers include Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Weights can be downloaded through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta lists support for engines such as vLLM, SGLang, llama.cpp and Unsloth.
Why it matters
For the past couple of years, Chinese developers have been the dominant force in open-weight models. Alibaba's Qwen, Moonshot AI's Kimi K3 and Z.ai's GLM family have been treated as the global frontrunners.
Meta now says its model competes at that level, and in some cases beats them. That is a vendor claim, so wait for independent benchmarks before anyone engraves anything.
The open-weight part is the point. Unlike closed models such as Claude or ChatGPT, developers can download the weights and modify them. That means fine-tuning, self-hosting and no per-token meter running.
"Runs at home" still needs a footnote. Two Mac Studios at the top of Apple's chip range is a desk setup, not a laptop. Cheaper than a data center, sure, but nobody is calling it a hobby budget.
The Apple tie-up is also odd. Meta says the model is optimized for M5 Ultra chips, which is a notable pairing between two companies not famous for sharing a toolbox.
What to do with it
- Check the benchmarks.Wait for independent evals comparing it with Qwen, Kimi K3 and GLM before trusting the "rivals" line.
- Try it via API first.OpenRouter, Together AI or Groq let you test it on your own coding tasks without buying hardware.
- Price the home setup.If you want local inference, cost out two M5 Ultra Mac Studios against API spend before getting excited.
- Read the license.Open weights can still carry terms, so check them before shipping anything commercial.
Ciao, sapiens. The frontier model now fits on your desk. Your desk, sadly, does not fit in your budget.


