Meta has published Muse Spark Open, an open-weights model with 816 billion parameters that the company says can be served from a pair of linked M5 Ultra Mac Studios using Apple’s MLX framework. The release landed on October 15 on Meta Superintelligence Labs’ blog, and Apple followed with a technical report detailing its own distributed-inference measurements — plus a strikingly cordial public exchange between the two companies’ chief executives.
- Huge model, no scorecard: 816B total parameters, 42B active per token, a one-million-token context window and natively multimodal input. Meta has not published a single benchmark figure of its own.
- Memory is the real constraint: more than 1.6 TB at full precision, roughly 493 GB compressed to 4 bits, which is why Apple spreads the weights across several machines.
- Apple’s measurements: 36.73 tokens per second across four M5 Ultra Mac Studios on Thunderbolt 5, with 4.8 seconds to first token at 8,192 tokens of context.
- The cheap route in: a single 512 GB M3 Ultra runs a 2-bit build at around 18 tokens per second.
A public thaw, without a contract
It is worth being precise about what this collaboration is not. Apple’s technical write-up never names Meta as the developer — the model surfaces only in file paths — and there is no joint press release, no disclosed financial terms and no exclusivity. What exists is engineering work plus a very visible gesture: Mark Zuckerberg thanked Apple on X, and John Ternus, Apple’s chief executive since September, answered in the same thread.
“It is amazing to run a model of the caliber and scale of Muse Spark Open with this efficiency, straight from your desktop.”
Meta’s AI chief, Alexandr Wang, said the two companies had worked together for months. The backdrop makes the exchange notable: Apple and Meta have been at odds since the 2021 iOS privacy changes, a shift Meta estimated cost it $10 billion in revenue in 2022.
Inside the architecture — and the missing evidence
Muse Spark Open is a mixture-of-experts transformer. It holds 816 billion parameters in total but activates only 42 billion per token, keeping the compute bill manageable. The context window reaches one million tokens, input is natively multimodal — text, images, audio and video — and output is text only. Meta positions it for coding and long-horizon agentic tasks, shipping it under its own Muse licence, which requires “built with Muse” attribution and a separate commercial agreement once a model-serving business passes $20 million in annual revenue. Weights sit on Hugging Face and in the Amazon, Google and Microsoft clouds, and the model works with the usual inference engines, from vLLM and SGLang to llama.cpp.
The gap is evidence. Meta calls it the most capable US open-weights model to date and points to an evaluation report that is not yet available. Independent scorer Artificial Analysis places it at 39 on its Intelligence Index — behind Chinese open models such as Mimo 2.6 Pro (46), GLM 5.3 (45), Kimi K3 (44) and Qwen 3.8 (40), and well below Meta’s own closed models Muse Wonder (52) and Muse Wonderful (56).
Desktop-sized, with an asterisk
At full precision the weights exceed 1.6 terabytes; a 4-bit version lands near 493 GB. Apple’s tests chained four M5 Ultra Mac Studios over Thunderbolt 5, reaching 36.73 tokens per second and 4.8 seconds to first token with 8,192 tokens of context — a wait that stretches to 40.6 seconds at 65,536 tokens. Its reference two-machine setup, each with 512 GB of unified memory, was clocked at roughly 25 tokens per second.
But the model card tells a more flexible story. With aggressive quantization, Muse Spark Open runs on a single 512 GB M3 Ultra: about 18 tokens per second from a 2-bit build occupying 304 GB, and up to 21 with a 3-bit MLX variant. “Two M5 Ultra Mac Studios” is therefore the promoted configuration, not a technical ceiling. Price and timing add friction — the M5 Ultra Mac Studio starts at $5,499, and the 512 GB configuration needed for the demo was not even on sale on announcement day, with Apple targeting late October and lead times expected to run into early next year. Cloud access is cheaper still, with third-party providers serving the model at $1.25 per million input tokens and $4.25 per million output tokens.
Why the gesture outweighs the product
The announcement matters most as a political signal. Two firms that spent five years trading public blows have just co-authored an engineering result, and both had reasons: Meta needs its open-weights bet to run somewhere other than a rack of Nvidia GPUs, while Apple gets to look central to the AI conversation without training a frontier model of its own. Set against that, a release with no published benchmarks, a restrictive licence and a hardware bill north of $10,000 for a configuration that ships later is not a product launch for ordinary users — it is a demonstration. Whether it becomes anything more will depend on the evaluation numbers Meta has promised, and on whether this détente survives its first disagreement.