Meta released the weights for Muse Spark Open on October 15, 2026, positioning the model as the most capable open-weight AI system ever released by a US company. According to Meta's official research blog, the model pairs a massive 816 billion total parameters with a sparse mixture-of-experts design that activates only 42 billion parameters per token, along with a 1 million-token context window.
The release lands one day after Meta launched Muse Wonder Full, its closed frontier model, and signals that the company intends to compete on both fronts: frontier performance through its API products and community accessibility through open weights. Meta has been the most consistent open-weight publisher among US frontier labs, and this release extends that strategy into the agentic era. For more context on this story, see our ongoing breaking AI news.
What Muse Spark Open Is
Muse Spark Open is a natively multimodal model built for coding and agentic work. Meta says it excels at long-horizon tasks, complex software engineering, and multi-step reasoning — the workloads that define agentic AI products in 2026. The 816B/42B activated-parameter split follows the sparse architecture pattern popularized by models like DeepSeek and Qwen, keeping inference costs manageable despite the enormous total parameter count.
The 1M-token context window places the model in the top tier for long-context work. Developers can feed entire code repositories, long document sets, or hours of transcripts into a single request without retrieval scaffolding.
The Apple Silicon Angle
The headline feature may be where the model runs. Meta says Muse Spark Open is the first model in its class to run on a desktop setup consisting of two Apple M5 Ultra Mac Studios using the latest MLX framework. Each M5 Ultra machine exposes up to 512 GB of unified memory to its integrated GPU, so two machines together hold the full model — effectively behaving, in Meta's words, like one supercomputer with a terabyte of unified memory. The chip's 1.2 TB/s memory bandwidth supplies the throughput the sparse architecture needs.
Meta demonstrated the setup running agentic workflows entirely on-device through Muse Code, including shipping real iOS changes to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve end to end — with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.
Availability and Ecosystem Support
The weights are downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Developers who prefer hosted inference can access the model through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
Meta also shipped optimizations across the major inference engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with support for AMD, Apple, Intel, and NVIDIA silicon. The open-weight release includes custom auto-tunable Helion GPU kernels in SGLang that deliver 1.03x to 3.95x kernel-level speedups on NVIDIA and AMD GPUs, and a 3 percent end-to-end speedup on AMD's MI350.
A Crowded Open-Weight Field
Meta's claim to the open-weight crown arrives in a market where Chinese labs have set much of the pace. Z.ai's GLM series, Alibaba's Qwen family, and DeepSeek's models have all shipped open weights at frontier-adjacent scale this year, and open-weight releases from US labs have been comparatively rare at the very top end. A US-origin model competing at this size matters for companies and governments that want strong models with weights they can inspect, audit, and modify without depending on a Chinese vendor or a closed API.
The timing is also notable. Meta launched Muse Wonder Full, its closed frontier model, just one day earlier. Shipping a closed flagship and an open counterpart within 24 hours lets Meta serve buyers who want turnkey API performance and the developers, researchers, and enterprises who want to self-host — two markets that increasingly demand different products.
Safety Posture
Open-weight releases at this scale carry obvious dual-use considerations. Meta says Muse Spark Open received additional safety post-training specifically for the open release, and that evaluations conducted under the company's Superintelligence Scaling Framework demonstrated strong safeguards across the risk categories measured. Independent red-teaming results will matter more than vendor claims, and the release will almost certainly be scrutinized by the same policy groups that have debated open-weight release norms throughout 2026.
What Desktop-Class Inference Changes
Running a model of this capability on two desktop machines rather than a rack of data-center GPUs has practical consequences beyond cost. Organizations with strict data governance requirements — hospitals, law firms, defense contractors, financial institutions — can keep prompts and documents entirely inside their own walls. Developers get latency and rate limits that depend on their own hardware rather than a provider's queue. And tinkerers get a frontier-scale model they can fine-tune, quantize, and modify without asking permission.
The trade-offs are real, too. Two M5 Ultra machines still represent a significant hardware investment, and a desktop cluster lacks the redundancy, uptime, and throughput of hosted inference. Meta's bet is that a meaningful slice of the market values ownership and privacy enough to accept those constraints — and that the demonstration effect of running an 816B-parameter model at home matters even for buyers who ultimately choose the API.
Why It Matters
The release sharpens the open-versus-closed debate that has defined the industry this year. Meta is betting that freely downloadable frontier-adjacent weights grow its ecosystem and undercut rivals that keep their best models proprietary. For developers, the practical significance is concrete: a top-tier agentic model that can be fine-tuned, self-hosted, air-gapped, and run on consumer-adjacent hardware rather than rented by the token.
Whether Muse Spark Open actually holds the "most capable US open-weight model" title depends on independent benchmarks, which were not available at publication time. But the combination of scale, long context, desktop-class hardware support, and broad inference-engine coverage makes it one of the most significant open-weight releases of the year — and a direct challenge to every closed model sold at premium API prices.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
