The short version
Muse Spark Open gives US and EU teams a strong, US-built open-weight option for coding agents that can run inside their own infrastructure. The 816B mixture-of-experts model activates 42B parameters per token, handles up to one million tokens of context, and is available as weights or through hosted APIs such as Together AI at $1.25 per million input tokens. For teams planning GenAI integration where code or data cannot leave their environment, it is worth a benchmark slot, with the license read first.
What did Meta release?
On October 15, Meta published the weights of Muse Spark Open, which it calls the most capable US open-weight model to date. According to Meta’s developer blog, it is the open-weights version of Muse Spark 1.3 and is aimed at “long-horizon agentic work” such as reading codebases, planning changes, calling tools and recovering when a step fails. Meta says it added extra safety post-training so the model holds up without the system-level guardrails of its hosted products.
The release has a clear strategic shape. Bloomberg framed it as Meta releasing an open model while keeping its best technology closed: three days earlier, Meta launched the proprietary Muse Wonder family, whose strongest version is not even on its API. CNBC highlighted the other half of the story, a partnership with Apple so developers can run the model locally on high-end Macs.
How open is the Muse Model License?
Open enough for most product teams, but with conditions. Muse Spark Open ships under the Muse Model License rather than Apache 2.0, which Meta used for its smaller Muse Glimmer model in August. VentureBeat reports that internal deployment, research and most products are covered, while businesses that serve the model to others and exceed $20 million in aggregate gross revenue need a written agreement with Meta. Attribution such as “Built with Muse Spark Open” is required, and Meta claims no ownership of generated outputs.
Access is gated too. VentureBeat notes that the Hugging Face download asks for a full legal name, date of birth and organization name. None of this blocks a pilot, but procurement and legal teams will want to sign off before the model touches customer data.
What does it take to run Muse Spark Open?
Memory, mostly. Meta and Apple describe a desktop setup of two M5 Ultra Mac Studios with 512GB of unified memory each, or four 256GB machines, running through Apple’s MLX framework; VentureBeat puts the four-machine option at about $43,000. In the data center, Hugging Face lists a single B200 node for the 4-bit NVFP4 checkpoint and four nodes of eight H200 GPUs for full BF16 precision. The model also runs on SGLang, vLLM, TensorRT-LLM and llama.cpp across NVIDIA, AMD, Intel and Apple hardware.
Most teams will start hosted. Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI for API access, and Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry for deploying the weights in your own cloud account. VentureBeat cites Together AI pricing of $1.25 per million input tokens and $4.25 per million output tokens.
How does Muse Spark Open compare with other open models?
It is competitive, not dominant. On the Artificial Analysis Intelligence Index cited by VentureBeat, Muse Spark Open scores 39, behind Chinese open models MiMo-V2.6-Pro at 46, GLM-5.3 at 45 and Kimi K3 at 44. Its strength is tool use: 89.2 on MCP-Atlas against 84.2 for GLM-5.3. It trails on terminal and repository-scale work, with 76.4 on Terminal-Bench 2.1 versus 85.0 for Kimi K3, and 60 on DeepSWE 1.1 versus 68.
For many US and EU buyers, though, the comparison that matters is not only quality. A US-built model with documented provenance can be easier to clear through vendor-risk reviews than an equivalent Chinese model, even when both run on your own servers.
What it means for US & EU software teams
The biggest change is for teams that could not use frontier-class coding agents because source code, health data or payment data cannot go to a third-party API. A self-hosted Muse Spark Open deployment in your own cloud region keeps prompts, repositories and outputs inside your perimeter, which simplifies GDPR data-transfer questions and sector rules such as HIPAA and PCI DSS. It does not remove them: logging, access control and retention still need to be designed.
Fine-tuning is the second opening. Meta supports supervised fine-tuning, LoRA and reinforcement learning through tools such as TorchTitan, NVIDIA NeMo, Hugging Face TRL, Fireworks and Together. That makes it realistic to adapt the model to an internal framework, codebase conventions or a domain-specific API, something closed frontier models rarely allow at this size.
The trade-off is operations. Running an 816B model means owning inference capacity, upgrades, evaluation and security patching. For a team without MLOps depth, a hosted endpoint with a private deployment or zero-retention terms may deliver most of the benefit at a fraction of the effort. And the gap between this open release and Meta’s closed Muse Wonder is a reminder to keep a provider-agnostic layer in front of any model.
How to evaluate Muse Spark Open this quarter
- Start hosted. Benchmark it through a hosted API against your current coding model on real tickets before buying hardware.
- Test where it claims strength. Measure multi-step tool calls and MCP workflows, and check its weaker spots in long terminal and repository tasks.
- Read the license early. Confirm whether your product counts as serving the model and whether the $20 million threshold applies.
- Price self-hosting honestly. Compare GPU or Mac Studio costs, power and engineering time with hosted per-token prices at your real volume.
- Plan for adaptation. If it passes, pilot a LoRA fine-tune on your own codebase conventions before a full rollout.
Frequently asked questions
What is Meta Muse Spark Open?
Muse Spark Open is an open-weight AI model Meta released on October 15, 2026. It is a mixture-of-experts model with about 816 billion total parameters and 42 billion active per token, a context window of up to one million tokens, and text, image, audio and video input. It shares a base with Muse Spark 1.3 and is tuned for coding and agent work.
Is Muse Spark Open free for commercial use?
Mostly, but it is not Apache 2.0. It ships under Meta's custom Muse Model License. VentureBeat reports that internal deployment and most products are covered, while model-as-a-service businesses above $20 million in aggregate gross revenue need a separate written agreement with Meta, and attribution such as “Built with Muse Spark Open” is required. Have legal review the license before production use.
What hardware does Muse Spark Open need?
A lot of memory. Meta and Apple describe a desktop setup of two M5 Ultra Mac Studios with 512GB of unified memory each, or four 256GB units, running via MLX. In the data center, Hugging Face lists a single B200 node for the NVFP4 version and four nodes of eight H200 GPUs for full BF16 precision. Most teams will start on a hosted API instead.
How does Muse Spark Open compare with Chinese open-weight models?
It leads on some agent benchmarks and trails on others. VentureBeat reports an Artificial Analysis Intelligence Index score of 39, below MiMo-V2.6-Pro at 46, GLM-5.3 at 45 and Kimi K3 at 44. It posts 89.2 on MCP-Atlas tool use versus 84.2 for GLM-5.3, but 76.4 on Terminal-Bench 2.1 versus 85.0 for Kimi K3.
Sources
Meta — Build with Muse Spark Open (October 15, 2026)
Hugging Face — Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta (October 15, 2026)
TechCrunch — Meta launches an open-weight AI model that runs on Macs (October 15, 2026)
Bloomberg — Meta Releases Open AI Model While Keeping Its Best Tech Closed (October 15, 2026)
CNBC — Apple and Meta team up to make Muse AI model run on Mac computers (October 15, 2026)
VentureBeat — Meta unveils Muse Spark Open, the new top U.S. open weights model (October 2026)