The short answer: On October 15, 2026, Meta Superintelligence Labs released Muse Spark Open, an 816-billion-parameter open-weight AI model designed for coding and agentic work. It runs with only 42 billion parameters active per token, supports a 1-million-token context window, and was available on seven hosted inference providers and four weight-distribution channels on day one. Meta calls it the most capable US open-weight model to date, and Meta Chief AI Officer Alexandr Wang has made clear the company’s very best models will stay closed. The model went live October 15, 2026.
What Meta Announced on October 15
Meta pushed Muse Spark Open into the wild on October 15, 2026, through Meta’s developer blog, and the rollout looked less like a research paper drop and more like a coordinated product launch. The model went live simultaneously across seven hosted-inference providers and four separate weight-distribution channels, with infrastructure partners lined up before the announcement even went public.
Muse Spark Open is natively multimodal on the input side: it takes text, images, video, and audio, and produces text output. Meta is positioning it squarely at coding and agentic workloads rather than consumer chit-chat, with long-horizon tasks, complex software engineering, and multi-step reasoning named as its primary use cases. That framing puts it in direct competition with the current generation of agentic coding models, not with chatbots.
For developers, the important context is that this is an open-weight release. Anyone can download the weights, inspect them, fine-tune them, and deploy them on their own infrastructure. That is a very different proposition from Meta’s closed commercial offerings, and it is deliberate.
Muse Spark Open Specs at a Glance
Before getting into the details, here are the headline numbers Meta has published:
- Total parameters: 816 billion
- Active parameters per token: 42 billion
- Architecture: Mixture of Experts (MoE), 256 experts per layer with 8 selected per token
- Context window: 1 million tokens
- Inputs: Text, images, video, audio (multimodal input)
- Output: Text
- Release date: October 15, 2026
- License type: Open-weight (weights publicly downloadable)
How the Architecture Works
Mixture of Experts: 816 Billion Parameters, 42 Billion at a Time
The core design idea is the mixture-of-experts transformer. All 816 billion parameters are split across 256 experts per layer, and a router picks 8 of them for every token that moves through the network. The active parameter count, 42 billion per token, is the number that actually determines inference speed and GPU memory bandwidth during generation.
That matters because it keeps inference costs well below what a dense model of equivalent scale would require. It is the same basic design philosophy that has powered the largest open releases out of Chinese labs over the past 18 months. The difference this time is whose name is on the model card.
A 1-Million-Token Context Window
The 1-million-token context window is large enough to hold entire codebases, long technical specifications, or multi-document research bundles in a single pass. A coding agent that has to re-read a 40-file repository on every turn burns tokens and time; one that can hold the whole thing in context does not.
Meta’s framing around long-horizon tasks suggests Muse Spark Open is built to stay coherent across many sequential tool calls, not just to answer a single prompt well. For teams building agents that plan, execute, check their work, and revise, that coherence across long sessions is exactly the property that matters most.
Native Multimodal Input
Being natively multimodal on the input side, handling text, images, video, and audio, while producing only text output is a deliberate scoping choice. Meta is not trying to compete with image or video generation models here. The model is built to look at a screenshot of a bug, read a video of a reproduction case, or parse a voice note describing a feature request, and then write code or reasoning in response. For engineering workflows, that input mix covers most of what a developer would show a senior colleague when asking for help.
Where You Can Run It From Day One
The part of this launch that deserves more attention than it has gotten is the sheer breadth of day-one availability. Most open-weight releases drop weights on Hugging Face and wait for the community and infrastructure providers to catch up over the following weeks. Meta instead lined up eleven distribution partners before the public announcement even went out.
Hosted Inference: Seven Providers
If you just want an API and do not want to manage GPUs, Muse Spark Open is served through hosted inference on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. That means you can start calling the model through an API key on day one, and you can comparison-shop providers on price, latency, and features instead of being locked to a single host.
Download the Weights: Four Channels
If you want to self-host, fine-tune, or inspect the model, the raw weights are downloadable through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. Microsoft confirmed its side of the arrangement through its own Azure AI Foundry blog, and Intel announced day-0 support for the model through its newsroom, a signal that chipmakers are now racing to validate new frontier-scale open releases the moment they ship rather than weeks later.
This is a coordinated enterprise sales motion sitting behind what looks, on the surface, like a research release. When Amazon, Google, Microsoft, and Hugging Face all agree to host the same model’s weights on day one, that represents a meaningful amount of engineering and business-development coordination across companies that compete with each other in other contexts.
Running an 816-Billion-Parameter Model on Two Mac Studios
One of the most talked-about details from the launch is Meta’s demo of Muse Spark Open running on a desktop setup built from two Apple M5 Ultra Mac Studios using Apple’s MLX framework. For a model carrying 816 billion total parameters, that is a notable claim. Frontier-scale open models of this size have typically been framed around data-center GPU clusters, not workstation hardware sitting under a desk.
If the claim holds up under independent testing, the practical effect is that individual developers and small engineering teams could run a model with capabilities Meta is positioning near the frontier without paying for cloud GPU time at all. That changes the calculus for anyone weighing a subscription to a closed, hosted model against the upfront cost of hardware that pays for itself over time. It also reinforces Apple’s pitch that unified memory architecture on Mac Studio hardware is a legitimate platform for serious local AI work, not just a convenience for running smaller, quantized models.
The local-hardware angle resonated beyond the US tech press. South Korea’s Chosun Ilbo covered the launch as a joint Meta-Apple moment, framing the unveiling around both companies’ involvement. Apple separately signaled that the open-source collaboration could support new on-device AI features across its ecosystem, which it hopes may encourage users to consider upgrading to newer iPhone models.
Meta has been expanding its own enterprise AI footprint this month, so this open release lands alongside the company’s commercial push. If you want the background on Meta’s business-facing AI products, see our earlier look at Meta Muse for small business.
Muse Spark Open vs Muse Wonder: Meta’s Two-Tier AI Strategy
Muse Spark Open is not Meta’s most capable model. That distinction reportedly still belongs to Muse Wonder, the closed, higher-tier system Meta has positioned against OpenAI’s commercial offerings. The split mirrors a strategy Meta’s own leadership has described openly: ship a strong open-weight model to build developer goodwill and ecosystem lock-in, while keeping the most advanced capabilities behind a closed, monetized product.
Alexandr Wang, Meta’s Chief AI Officer, has articulated that logic directly, saying that because of both safety and capabilities, Meta will generally not open-source its most powerful models. That draws a clean line around Muse Spark Open’s role in Meta’s lineup: it is the company’s strongest card in the open-weight game, but it is deliberately not the best card in Meta’s full hand.
That two-tier approach is becoming the default playbook across the industry. A closed flagship justifies premium pricing and keeps a company’s absolute best capabilities proprietary, while a strong open release builds the kind of developer trust, third-party tooling, and fine-tuning ecosystem that a closed-only strategy cannot. Meta is betting that Muse Spark Open’s reach across seven inference providers and four weight-hosting platforms builds exactly that ecosystem faster than a closed-only approach ever could.
Muse Spark Open vs DeepSeek, Qwen and GLM
Context matters here. For most of 2025 and into 2026, the open-weight leaderboard conversation has been dominated by Chinese labs. DeepSeek’s mixture-of-experts releases forced a rethink of what efficient at scale could mean for open models. Alibaba’s Qwen line built out one of the broadest multilingual and multimodal open catalogs in the industry. Zhipu AI’s GLM models carved out a strong reputation specifically in reasoning and agentic tool use. Meta’s own Llama family, meanwhile, had started to look like the previous generation’s open-weight champion rather than the current one.
Muse Spark Open is Meta’s answer to that shift, and the most capable US open-weight model to date framing is a direct claim about reclaiming ground from US-based competitors rather than conceding the open category entirely to labs operating under different regulatory and export conditions. It is worth being precise about what is and is not established here: Meta has not published independently reproduced benchmark comparisons against DeepSeek, Qwen, or GLM alongside this release, so any specific performance ranking should be treated as a claim to watch, not a confirmed fact.
For more on how the open-weight race has been playing out this year, see our coverage of Mistral Large 4, Europe’s major open-weight release from earlier this month.
Why the Distribution Strategy Matters More Than the Specs
The appetite for open-weight models among enterprise buyers has been building for a while, which helps explain why Meta lined up eleven distribution partners for this launch instead of a quiet Hugging Face drop. Research from McKinsey and Company found that more than 50 percent of respondents reported leveraging open source in each of the data, models, and tools areas of the technology stack, which shows open-weight adoption is no longer a fringe engineering preference but a mainstream procurement pattern.
The same research found that more than 70 percent of respondents said they were open to either open-source or proprietary technologies across areas of the technology stack, which suggests most enterprise technology buyers are pragmatic rather than ideological about open versus closed models. That favors whichever lab ships the most accessible, best-distributed option at a given moment. Notably, McKinsey also found that among respondents using open-source models, 61 percent had used Meta’s Llama, compared with 40 percent who had used Google’s Gemma. That gives Meta a real incumbency advantage heading into the Muse Spark Open launch: a large share of the enterprise open-model user base already has existing Llama tooling and pipelines that could plausibly carry over.
The timing also fits a broader industry pattern. Labs are racing to ship agents that can do real work, not just chat, and the infrastructure behind those agents matters as much as the models. Our recent coverage of Google’s PageBreak AI agent, which autonomously found hundreds of real vulnerabilities, shows the same shift toward models evaluated on what they can accomplish over long sessions rather than on single answers.
Price and Availability
Muse Spark Open is free to download and use under its open-weight release, which is the whole point of the open strategy. If you self-host, your cost is your own compute. If you use one of the seven hosted inference providers, you pay that provider’s standard token pricing, which varies by vendor.
There is no sign-up waitlist and no staged rollout. The model was available on all eleven distribution channels at launch on October 15, 2026. For hardware, the official story is GPUs, whether cloud or self-hosted, with the two-Mac-Studio MLX demo as the intriguing local-run alternative. The chip landscape behind all of this keeps expanding too; Microsoft just announced its own Maia 300 AI chip, aimed squarely at cutting inference costs for exactly this kind of workload.
What This Means for You
If you are a developer or a small team, the practical takeaways are straightforward. First, you now have a frontier-positioned open model you can run without a Meta relationship, which removes a layer of vendor dependence from any agent or coding-assistant project. Second, the eleven-channel distribution means you can test it on hosted APIs today and move to self-hosting later without changing models. Third, the 1-million-token context and multimodal input are genuinely useful for the exact workloads where open models have historically lagged closed ones: long coding sessions with big repositories.
The open question, and the honest one, is benchmarks. Until independent researchers reproduce head-to-head comparisons against DeepSeek, Qwen, and GLM, treat the most capable US open-weight claim as marketing to verify rather than fact. The specs are real and the distribution is real. The performance crown is still to be decided.
Frequently Asked Questions
When was Meta Muse Spark Open released?
Muse Spark Open was released on October 15, 2026, by Meta Superintelligence Labs. It was available on all seven hosted inference providers and all four weight-distribution channels on day one.
Is Meta Muse Spark Open free?
The model weights are free to download under the open-weight release. If you self-host, you pay only for your own compute. If you use a hosted inference provider such as Fireworks AI, Groq, or Together AI, you pay that provider’s normal API pricing.
How big is Muse Spark Open?
Muse Spark Open has 816 billion total parameters with 42 billion active per token, using a mixture-of-experts design with 256 experts per layer and 8 selected per token. It supports a 1-million-token context window and natively multimodal input covering text, images, video, and audio.
Can I run Muse Spark Open on my own computer?
Meta demonstrated the model running on two Apple M5 Ultra Mac Studios using Apple’s MLX framework, which suggests serious local deployment is possible on high-end workstation hardware. Independent testing has not yet confirmed real-world performance or quantization details, so treat the demo as promising rather than proven.
Is Muse Spark Open better than DeepSeek or Qwen?
Meta claims it is the most capable US open-weight model to date, but the company has not published independently reproduced benchmarks against DeepSeek, Qwen, or GLM with this release. Wait for independent benchmark results before treating any ranking as confirmed.
What is the difference between Muse Spark Open and Muse Wonder?
Muse Spark Open is Meta’s open-weight model, free to download and self-host. Muse Wonder is Meta’s closed, higher-tier system positioned against OpenAI’s commercial offerings. Meta’s Chief AI Officer Alexandr Wang has said the company will generally not open-source its most powerful models, so the best capabilities stay behind the closed product.