Meta Opens Muse Spark: 816B Parameters, the Most Capable US Open-Weight Model

Meta released Muse Spark Open on October 15, two days after the Muse Wonder launch, and the claim on the page is the story: “the most capable US open-weight model to date.” The specs: 816B total parameters with 42B active in a sparse MoE, natively multimodal, a 1M-token context window, built for coding, agentic work, long-horizon tasks, and multi-step reasoning. Weights are downloadable from Hugging Face, Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry under the “Muse Model License with limited commercial terms,” which allows many developers to use it free. Two days after Muse Wonder, Meta now has both a hosted agent family and an open-weights flagship in the same week.

The Desktop Demo Is the Flex That Matters

The distributed-requirements story is where Muse Spark Open breaks new ground: Meta says it is the first model in its class to run fully on a desktop, using two Apple M5 Ultra Mac Studios (512GB unified memory each, 1.2TB/s bandwidth) via MLX. A trillion-parameter-class open model running on two desks is a concrete answer to the question of who can audit and build on it: not just clouds with A100 clusters, but well-funded labs, universities, and serious hobbyists. The optimization ecosystem shipped alongside: custom Helion GPU kernels in SGLang delivering 1.03 to 3.95x kernel-level speedups on NVIDIA and AMD GPUs, a 3% end-to-end speedup on AMD MI350, and day-one support across SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth, optimized for AMD, Apple, Intel, and NVIDIA hardware. Fine-tuning runs through PyTorch TorchTitan.

The Benchmarks Are a Link, Not a Table

The launch page has a Key Benchmarks section that links to an evaluation methodology report rather than printing numbers, which is a notable choice for a model claiming the US open-weight crown. The demos Meta did publish are concrete: the model shipped real iOS changes to the Ice Cubes Mastodon client and drove DaVinci Resolve end to end, reviewing footage frame by frame and producing vertical crops. It passed Meta’s internal Superintelligence Scaling Framework thresholds, with additional safety post-training applied before release. Independent evaluation of the 816B model against DeepSeek 4.1 Flash, Kimi K3, and the 5.5 family is the missing table, and every claim on the page waits on it.

The Open-Weight Race Gets Its American Contender

The open-weights story this month has been: DeepSeek 4.1 Flash at effectively-free pricing, Kimi K3, MiMo’s AutomationBench dominance, and Kolibri for Europe. Muse Spark Open is the first American entry at this scale since that wave started, and the license is the tell: “limited commercial terms” is not Apache 2.0, so Meta is open enough to claim the crown and closed enough to keep commercial leverage. That is the same line Kolibri walked (open weights, paid services) and the line Qwen Image 2.1 tripped over (research-only license stalled adoption). The market now has three licensing philosophies competing: Apache 2.0 (Liquid, Beam), limited-commercial (Meta, Qwen), and API-only (OpenAI, Anthropic).

What to Watch

Three things. First, independent benchmarks against DeepSeek 4.1 and Kimi K3: the “most capable US open-weight” title is a claim about relative standing, and the eval methodology report is where the numbers will surface. Second, the two-Mac-Studio deployment: if the 816B model genuinely runs useful workloads on 1TB of unified consumer-adjacent memory, the local-AI ceiling moves again within a week of ds4 and Strata pushing it from below. Third, the license in practice: what “limited commercial terms” permits, and whether developers building on it get burned the way Qwen Image 2.1’s research-only license did. Meta has one week of goodwill and a benchmark report to publish.