An 816B open-weight chip to play
On October 15, Meta released the weights of Muse Spark Open. It is a natively multimodal sparse-expert model: 816 billion total parameters, 42 billion active per token, a 1M-token context, ingesting text, image, video, and audio and emitting text. The license is Meta's own Muse Model License with limited commercial terms, and weights are on Hugging Face at meta-models/Muse-Spark-Open-816B-A42B. Its brief is clear: long-horizon, multi-step coding and agentic work.
Running it at home is a demo
Meta made noise about two Apple M5 Ultra Mac Studios running it via MLX, which sounds lovely, but do not be misled. In NVFP4 it needs roughly 700GB of accelerator memory, not something a laptop handles. Those two Mac Studios each max at 512GB unified memory, together just enough to hold it. Most teams will land on GPU clusters or hosted partners. The MLX path is for local residency experiments; do not promise customers it runs on a MacBook.
Three ways to get it
There are two ready paths. Self-host by downloading weights; SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth all support it at launch across NVIDIA, AMD, Intel, and Apple silicon. Or use a hosted API: Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together all serve it. The closed Wonder and Wonder Full are a separate line: Wonder goes through the Meta Model API, Wonder Full only with a Muse Code subscription, stronger but not on the public API tier.
Its place in enterprise self-hosting
Open weights reopen the self-host and sovereign-cloud options, right after the closed Argon and Wonder launches. For data-residency or EU-procurement scenarios, Spark Open gives more say than a purely API-bound closed model. But review the commercial terms on day one: the Muse Model License limits commercial use, so do not audit after a successful pilot. On FinOps, put partner token prices and GPU amortization on the same sheet.
Open or closed
Artificial Analysis gave Intelligence Index scores of 56 for Wonder Full, 52 for Wonder, and 39 for Spark Open. Lower score does not mean useless. Internal assistants, document processing, and customer-facing features often do not need the strongest model, they need accuracy at a sustainable cost. Open weight means control and customization, not a leaderboard top spot. Before promoting it, run three evals, long refactor, multimodal ticket, MCP tool loop, and promote only where it wins on cost per completed task.
How to choose against Wonder
Closed Wonder goes through the Meta Model API, and Wonder Full only with a Muse Code subscription; they are the ready-to-use, stronger, but token- or subscription-priced class. Spark Open is the download-weights, self-controlled, billed on your own hardware or hosting-partner class. Simply: want the fastest start and do not care about deployment, pick Wonder; mind data leaving the region, want to modify the model, want to amortize cost onto hardware long-term, pick Spark Open. They are not replacements, they are two different ways to own.
The truth about local runs
The two Mac Studios can run it demo makes people think consumer gear suffices, but those M5 Ultra machines at 512GB unified memory each are enterprise config, not your study MacBook. The 700GB VRAM need under NVFP4 means most people rely on cloud GPUs or hosting partners. The MLX path's value is a local-residency compliance experiment proving data can stay in-domain, not a toy a normal developer casually runs.
What open weights mean
Open weight's point is not free, it is control. You can fine-tune, post-train, drop the model into your private cloud, and audit what it actually does. For regulated industries that beats a leaderboard top score. The cost is owning deployment, updates, security, and performance, Meta carries those on the closed side. So choosing open is not for convenience, it is for sovereignty; teams without ops capacity are steadier on a hosted API.
Do not self-host first
Seeing open weights tempts people to deploy immediately; hold that urge. Self-hosting Spark Open assumes 700GB of VRAM or a solid hosting partner, and that two-Mac-Studio demo is enterprise top config, not your study machine. For most teams, validate on a hosted API like Baseten, Fireworks, or Together first, lower cost, faster start, no GPU to babysit, some with free tiers. Only after it proves it carries your tasks, and only if you have a hard residency requirement, does downloading weights and standing up an inference stack pay off. Starting with self-host risks getting stuck on VRAM, quantization format, and engine tuning, things unrelated to whether the model is good, wasting two weeks while never actually validating the model.
What this means for youIf your team already lives in the Hugging Face ecosystem and has a hard data-stays-in-domain rule, Spark Open is one of the few options pairing US open weights with a 1M context. But do not let the 816B size intimidate or tempt you. First ask whether you have 700GB of VRAM or a solid hosting partner, then decide between self-hosting and API. For most small teams, trying the hosted API first is the better value.