Meta's Muse Spark Open: 816 Billion Parameters, Free to Download

By Sophie Lindqvist

Meta has released Muse Spark Open, an open-weight AI model with 816 billion parameters that handles text, images, video, and audio in a single system. According to The Decoder, the model processes up to one million tokens of context and performs roughly on par with the most capable open-weight Chinese AI models on standard benchmarks.

That last detail is the story. Meta is directly targeting the Chinese models — Alibaba's Qwen, Zhipu's GLM, and Moonshot's Kimi — that have dominated the open-weight leaderboard in recent release cycles.

What is Muse Spark Open?

The model is open-weight, which means Meta publishes the trained parameters and lets developers download and run the system on their own hardware. That's distinct from an API-only model, where the weights stay locked inside a provider's data center.

The headline specifications, as reported by The Decoder:

• 816 billion parameters — large even by current open-weight standards
• Multimodal by default — text, images, video, and audio in one model
• Up to one million tokens of context — enough to process very long documents or extended video and audio in a single pass

In plain terms: this is a single system that can read, watch, and listen, at a scale that until recently only the largest proprietary models reached.

Why does the one-million-token context matter?

Context length sets how much material a model can consider at once. A one-million-token window lets a developer feed in entire codebases, hours of video, or long audio recordings without splitting the input into chunks.

For a multimodal model, the constraint bites harder. Video and audio consume tokens far faster than plain text, so a long context window is a practical requirement rather than a marketing number.

The takeaway: the context spec makes the multimodal claims usable, not just quotable.

How does it compare to Qwen, GLM, and Kimi?

The Decoder reports that Muse Spark Open performs roughly on par with the most capable open-weight Chinese models on benchmarks. That framing matters for the market.

Chinese labs have led the open-weight segment recently, with Qwen, GLM, and Kimi releases setting the reference point that developers compare against. If a Western lab matches that tier, it changes the calculus for companies that want open weights but prefer non-Chinese supply chains for policy or procurement reasons.

The benchmark parity claim is approximate — "roughly on par" — so buyers should expect wins and losses depending on the specific task.

In short: Meta is back in contention at the top of the open-weight table, not clearly ahead of it.

Why Apple's M5 Ultra is part of the story?

Muse Spark Open is optimized to run locally on Apple's M5 Ultra, The Decoder reports. That's an unusual detail for a model of this size.

An 816-billion-parameter model typically demands multi-GPU server setups. Optimizing for a single high-end workstation chip signals that Meta designed the model for local, on-device deployment — developers and researchers running it without cloud infrastructure.

Local execution also fits the open-weight pitch: if you can run the model yourself, you control the data, the latency, and the cost.

The takeaway: the hardware target tells you who Meta expects to actually use this thing.

What's at stake in the open-weight race?

Open-weight releases have become the competitive middle ground of the AI market. They don't generate direct revenue the way APIs do, but they shape developer ecosystems, set research agendas, and function as capability signaling between labs.

For Meta, a competitive open-weight model is leverage: it pulls developers away from exclusively API-dependent workflows and keeps the company relevant in the segment where Chinese labs have taken the lead.

For developers, the practical effect is more choice at the top tier — same rough capability level, different licensing jurisdictions and hardware targets.

The open question is how Muse Spark Open performs outside benchmarks, in real deployments. Based on what's been reported so far, the model lands Meta squarely back in the open-weight race — and the next Qwen, GLM, or Kimi release will tell us how long that lasts.