AI Revolution · Explainer · Updated Oct 2026
Meta's Muse Spark Open: frontier weights you can actually download
By James Brown · Published · Updated
On Oct 15 Meta released an 816B-parameter open-weight multimodal model aimed at coding and agents. What it is, what "open weight" buys you, and what it still costs to run.
On October 15, 2026, Meta released the weights for Muse Spark Open, calling it the most capable US open-weight model to date. Reuters covered the same release the same day. For builders watching the AI revolution, this is the interesting kind of news: not another closed demo, but parameters you can download, host, and adapt.
What shipped
According to Meta's research blog and the Hugging Face model card:
- 816 billion total parameters, with about 42 billion active per token (a mixture-of-experts design).
- A 1 million-token context window at inference.
- Natively multimodal inputs: text, image, video and audio, with text output.
- Built for coding and agentic work: long-horizon tasks, complex software engineering and multi-step reasoning.
- Released under the Muse Model License with limited commercial terms that Meta says allow many developers to use it for free.
Hugging Face's companion post adds architecture detail: roughly 256 experts per MoE layer with 8 selected per token, and the same tokenizer family as Muse Glimmer (Meta's smaller August open-weight release).
Where you can get it
You do not have to self-host to try it. Meta lists API access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Weights are posted on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. Supported runtimes include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth.
The catch: hardware is still the bottleneck
Open weights are not the same as "runs on your laptop." Meta's docs say the BF16 checkpoint is about 1.65 TB of weights; the NVFP4 checkpoint is about 537 GB, and that footprint excludes KV cache and runtime overhead. Meta's headline desktop path is two Apple M5 Ultra Mac Studios with MLX, using unified memory so the pair can hold the model. That is a serious machine, not a hobby pi.
So the story is less "everyone runs frontier AI at home tomorrow" and more "the ceiling of what a determined team can own just jumped." API partners cover the rest.
Safety, still Meta's framing
Meta says the open-weight release got extra safety post-training because it will ship without a company's live guardrails, and that it met thresholds in Meta's Superintelligence Scaling Framework. Treat that as the publisher's claim: useful context, not an independent audit. Open weights raise the usual dual-use questions; that is why the model card and license matter.
Why this matters for the revolution
Closed models set the pace of demos. Open weights set the pace of ecosystems: fine-tunes, agents that keep data on-prem, research reproducibility, and price pressure on API vendors. Zuckerberg's public line, as Reuters notes, is that advanced AI should not stay concentrated in a few companies. Whether you buy the rhetoric or not, downloading Muse Spark Open is a concrete step in that direction.
- If you build agents: a strong open coding/agent model is another option when you need control, offline use, or custom post-training.
- If you invest in the theme: Meta (META) sits on both sides of the ledger: huge closed products and open releases that shape the developer stack. Track the company on our Markets snapshot; do not treat a model launch as a buy signal.
- If you are learning: read the model card, try an API partner first, and only then think about self-hosting cost.
Sources
- Meta AI Research, Introducing Muse Spark Open (Oct 15, 2026)
- Hugging Face model card, Muse-Spark-Open-816B-A42B
- Hugging Face blog, Welcome Muse Spark Open (Oct 15, 2026)
- Meta Model API docs, Get the model (Muse Spark Open)
- Reuters, Meta releases open-weight Muse Spark model (Oct 15, 2026)
More AI explainers See the market snapshot →
FAQ
What is Muse Spark Open?
Meta's October 15, 2026 open-weight multimodal model with about 816B total parameters (42B active), a 1M-token context, and a focus on coding and agentic tasks.
Is it free to use?
Meta released the weights under the Muse Model License with limited commercial terms that it says allow many developers to use it free. Hosting and API inference still cost money and hardware.
Can I run it on a normal PC?
Not realistically. Meta's docs put the compressed NVFP4 weights around 537 GB and the BF16 weights around 1.65 TB before runtime overhead. Meta demonstrates desktop use on two Apple M5 Ultra Mac Studios with MLX.
How is this different from ChatGPT-style APIs?
You (or a host) can download the parameters, run them where you choose, and fine-tune them. Closed chat products keep the weights private and sell access through their interface.