Meta's Biggest Open-Weight Model Is the Best One Built in America. Its Own Benchmark Table Explains the Qualifier

Meta Superintelligence Labs released Muse Spark Open on 15 October, and the place to start is the wording of Meta's own headline. The company calls it "the most capable US open-weight model to date." Not the most capable open-weight model. The most capable American one.

That qualifier is not modesty. It is the conclusion Meta's own benchmark table reaches.

What shipped

Muse Spark Open is 816 billion total parameters with 42 billion active per token — a mixture-of-experts transformer with 62 layers, 256 experts of which 8 fire per token, a hidden dimension of 8,192, grouped-query attention over 128 query and 16 key-value heads, and a 202,048-token vocabulary, fronted by a 1.8-billion-parameter Perception Encoder. Text, image, video and audio go in; text comes out; the context window is one million tokens.

The weights are on Hugging Face, mirrored into Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry, served by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and supported in SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth. That is a genuinely wide launch surface.

Open weights is not open source. The Muse Model License requires anything built on the model to display "Built with Muse Spark Open," states that the model "is not intended for individuals under the age of 18," and requires a licensee offering the model to third parties as a service to notify Meta and enter a separate written agreement once that service passes $20 million in gross revenue over any twelve consecutive months. Internal use, research and non-hosted products are outside that gate.

The table

Every figure below is from Meta's model card. No third party has reproduced any of it.

Meta leads on agentic tool use and on long-horizon job tasks. It trails Moonshot's Kimi K3 on computer use, software engineering and terminal work, trails Zhipu's GLM 5.3 on the aggregate intelligence index, and trails both Qwen and Kimi on graduate-level science questions. The companies in the right-hand column are all Chinese. Meta chose a geographic adjective because it had to.

Where Meta does lead by a wide margin is refusal behaviour. The card reports 99.2 percent jailbreak refusal on an internal cyber benchmark and 94.8 percent across aggregated public sets, against 75.5 and 70.2 percent for Kimi K3 in the same harness, plus 98.1 percent honesty on MASK, a zero percent misbehaviour rate on agentic misalignment, and 0.5 on StrongREJECT v2 where the other listed models score between 62.9 and 86.4. For a model whose weights anyone can download and post-train, that gap matters more than a point of SWE-bench — and Meta's own authors note the results hold only for the released checkpoint, and that modified versions may be less safe.

Apple did the hard part

The second half of the launch is the more surprising one. Apple published its own research post the same day, "Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio," describing how to run the model across linked machines using MLX, RDMA over Thunderbolt 5 — a low-latency path introduced in macOS 26.2 and exposed through a jaccl backend — and either pipeline or tensor parallelism.

The numbers are specific. Muse Spark Open needs more than 1.6 TB in BF16 and about 493 GB quantized to NVFP4, so it cannot fit on one machine. Two M5 Ultra Mac Studios decode at 15.74 tokens per second with pipeline parallelism and 24.99 with tensor parallelism; four reach 36.73, with time-to-first-token falling from 8.4 seconds to 4.8. Meta's line is that "two Mac Studios behave like one supercomputer with a terabyte of unified memory." Apple CEO John Ternus posted that it was "amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk."

The desk is the caveat. The M5 Ultra Mac Studio starts at $5,499 with 96 GB of memory; the 512 GB build-to-order configuration this setup requires has not shipped yet and is expected late this month, and the existing 256 GB machine already costs north of $10,000. A four-machine mesh is a rack with a nicer finish.

Two strategies, four days apart

Meta shipped its first closed frontier model, Muse Wonder, on 12 October, with no weights and its best variant locked behind a $200 subscription. Three days later it shipped its largest open-weight model with a $20 million revenue gate and a partner that spent fifteen years as its loudest antagonist on privacy. Meta's AI chief Alexandr Wang, asked about that, told Axios his team was "excited for mommy and daddy to be playing nice again."

Both releases are coherent once you read them as distribution plays rather than research ones. Wonder monetises the frontier. Spark Open commoditises the tier below it, keeps Meta in the conversation American labs have largely conceded to Hangzhou and Beijing, and does it on hardware a developer can buy rather than rent from a cloud. What it does not do is take the open-weight lead back. Meta's own table says so.

Sources: Meta AI Research: Introducing Muse Spark Open · Muse Spark Open model card on Hugging Face · Apple Machine Learning Research: Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio · 9to5Mac: Meta launches new Muse Spark Open model, optimized to run on Mac Studio · MacRumors: Meta and Apple Bring Muse Spark Open AI Model to Mac Studio · TechCrunch: Meta launches an open-weight model that runs on Macs · Techstrong.ai: Meta, Apple Partner to Bring Massive 816B-Parameter Muse Spark Open AI Model to Desktop Macs