Illustration of an open-source AI model: glowing neural network nodes merging with code brackets

Meta Releases Muse Spark Open: What the 816B Open-Weight AI Model Is and How to Try It

ByChris Wade

Key takeaway: On October 15, 2026, Meta released the open weights for Muse Spark Open, an 816-billion-parameter AI model with 42 billion active parameters, a 1-million-token context window, and native multimodal input. Meta calls it the most capable US open-weight model to date. You can download it free from Hugging Face, run it through hosted API providers, or run it locally — if you have about a terabyte of memory.

What Meta announced

Meta Platforms released the open weights of Muse Spark Open on October 15, 2026, making the model — previously proprietary since its April debut — free for developers to download, run, and build on. The release was reported by SiliconANGLE and 9to5Mac, both citing Meta’s official announcement.

Muse Spark Open is a mixture-of-experts (MoE) model: all 816 billion parameters exist, but only about 42 billion activate for any given token, which keeps inference costs far lower than the headline size suggests. It accepts text, images, audio, and video as input and produces text output, with a context window of up to 1 million tokens — enough to hold entire codebases or hours of video in a single conversation.

The model is based on Muse Spark 1.2, which shipped in August 2026. Meta added extra safety post-training so the model “holds up without the guardrails of Meta’s own services,” since anyone running the weights locally gets none of Meta’s built-in protections. Meta also said the model cleared the release thresholds in its internal Meta Superintelligence Scaling Framework, with a full evaluation report still to come.

Why it matters

This is a meaningful moment in the open-weight race. Meta’s previous open-weight release, Muse Glimmer, was a 30-billion-parameter model — Muse Spark Open is roughly 27 times larger. And Meta’s most capable systems, the Muse Wonder models introduced on October 12, remain closed, available only through the Meta Model API and Muse Code. The company is effectively giving developers a heavyweight model to run themselves while keeping its frontier tier behind an API.

The timing matters because Chinese open-weight models have been leading the coding benchmarks. In Meta’s own benchmark table, Muse Spark Open scored 61.8 on the JobBench agent test — more than double the 31.1 of Inkling, the US open model from Thinking Machines Lab — but on Terminal-Bench 2.1 coding, Chinese models led: Kimi K3 at 85, GLM 5.3 and Qwen3.8 both above Muse Spark Open’s 76.4. Where Muse Spark Open flipped the order was tool use: its 89.2 on MCP-Atlas, which tests plain-language tasks across 220 tools on 36 live servers, beat every rival Meta listed, Chinese models included.

The surprising hardware story: two Mac Studios

Perhaps the most unusual part of the announcement is where Meta says the model can run. Meta tested Muse Spark Open on a pair of Apple Mac Studio desktops with M5 Ultra chips, linked through Apple’s MLX framework, giving roughly a terabyte of unified memory between them. On that setup, Muse Code — Meta’s coding tool — shipped real iOS changes to Ice Cubes, an open-source Mastodon client, including bug fixes, new features, and a full build. The same setup also drove DaVinci Resolve end to end, reviewing footage frame by frame to find the strongest moments and crop them vertically.

That demo is a signal: an advanced coding-and-agent model that can run on a desktop rather than a data center cluster. The catch is that the 512 GB Mac Studio configuration Meta used is not yet widely available, and two machines still add up to roughly $20,000 of hardware. For most people, this is a proof of concept, not a weekend project.

How to try it (and the license fine print)

The weights are free to download from Hugging Face and from the model catalogs of AWS, Google Cloud, and Microsoft. At launch, seven inference providers offer hosted access, including Groq and Fireworks AI — the fastest route for developers, requiring only an account and API key.

The license is free but not open source. Meta’s new Muse Model License applies “limited commercial terms”: developers pay no fee to adapt the model or ship it inside their software, but companies with over $20 million in gross revenue in 12 months that sell the model as a hosted service need a separate agreement with Meta. Products built on it must display a “Built with Muse Spark Open” label, and military and critical-infrastructure uses are banned.

Analysis: where the claims need independent testing

Analysis (not a verified fact): Every performance number in this article comes from Meta or its launch partners — no independent benchmarks have been published yet, so treat the “most capable US open-weight model” claim as Meta’s framing, not an established result. The honest read: Muse Spark Open looks strongest at agentic tool use and long-context work, and mid-pack among large open models on raw coding. If you build agentic workflows and want open weights with a million-token window, it’s worth evaluating. If raw coding power is your priority, the Chinese open models currently score higher on Meta’s own table.

Frequently asked questions

What is Meta Muse Spark Open?

Muse Spark Open is the open-weight version of Meta’s Muse family, released October 15, 2026. It’s an 816-billion-parameter mixture-of-experts AI model (42 billion parameters active per token) with a 1-million-token context window and native multimodal input — text, images, audio, and video in, text out. It’s designed for coding and long-running agent work, and its weights are free to download.

Can I run Muse Spark Open on my own computer?

Only with very serious hardware. Meta’s reference setup is two Apple Mac Studios with M5 Ultra chips and about 1 TB of combined unified memory — roughly $20,000 of hardware, and the 512 GB configuration Meta used isn’t widely available yet. For everyone else, the practical route is a hosted API from one of the seven launch providers, such as Groq or Fireworks AI.

Is Muse Spark Open free to use commercially?

The weights are free to download, and there’s no fee to adapt the model or ship it inside your software. But the license requires a “Built with Muse Spark Open” label on products, companies with over $20 million in annual gross revenue that sell it as a hosted service need a separate agreement with Meta, and military or critical-infrastructure uses are banned. It’s free, but it’s not open source.

How does Muse Spark Open compare to Chinese open models like Kimi K3?

On coding, Chinese models lead: in Meta’s own benchmark table, Kimi K3 scored 85 on Terminal-Bench 2.1 versus 76.4 for Muse Spark Open, with GLM 5.3 and Qwen3.8 also ahead. On agentic tool use, Muse Spark Open leads with an 89.2 on MCP-Atlas, beating every rival Meta listed. All of these numbers are vendor-reported; no independent benchmarks have been published yet.

What is the difference between Muse Spark Open and Muse Wonder?

Muse Spark Open is the open-weight release of the Muse Spark model family — you can download and run it yourself. Muse Wonder, launched October 12, 2026, is Meta’s closed frontier model, available only through the Meta Model API and Muse Code. Meta keeps its most capable systems closed while giving developers a smaller-but-still-large open model to run locally.

Sources


Similar Posts