Published: 16 October 2026
Meta released Muse Spark Open on 15 October 2026, an open-weight model that anyone can download, run on their own hardware and fine-tune for free. Meta calls it the most capable American open-weight model, and its own benchmark table supports that. The same table shows China’s GLM 5.3 and Kimi K3 scoring higher overall and on coding, while Muse Spark Open leads when an agent has to use tools.
The model comes from Meta Superintelligence Labs. It has 816 billion parameters, of which only 42 billion work at any moment. This design, called mixture of experts, makes it faster than its size suggests. It reads text, images, audio and video, answers in text, and accepts up to one million tokens (pieces of words) in a single prompt. Meta built it for coding and for agents: software that carries out multi-step tasks on its own.
Muse Spark Open is not Meta’s best model. Muse Wonder, launched on 12 October, stays closed. “Because of both safety and capabilities, we will generally not open-source our most powerful models,” Meta’s chief AI officer Alexandr Wang said, as reported by The Next Web.
Muse Spark Open leads on tool use and trails GLM 5.3 and Kimi K3 on coding
Meta compared its model with the open models developers actually choose between, and the result is mixed.
| What is measured | Muse Spark Open | GLM 5.3 | Kimi K3 |
|---|---|---|---|
| Overall score (Artificial Analysis Intelligence Index) | 39 | 45 | 44 |
| Using tools through MCP (MCP-Atlas) | 89.2 | 84.2 | 82.3 |
| Agent tasks (JobBench) | 61.8 | 61.4 | 54.3 |
| Coding in a terminal (Terminal-Bench 2.1) | 76.4 | 83.9 | 85.0 |
| Software engineering (DeepSWE 1.1) | 60 | 67 | 68 |
| Hard questions without tools (HLE-Diamond) | 26.1 | 16.4 | 22.2 |
Meta measured all these figures itself, at the model’s highest reasoning setting. VentureBeat found no independent Artificial Analysis score yet, and Meta’s full evaluation report is still to come. Read a gap of one or two points as a tie.
MCP is the common plug that connects AI agents to other software. A high MCP-Atlas score matters most if you build agents that search a mailbox, query a database or file a ticket. For writing and fixing code, GLM 5.3 and Kimi K3 remain the stronger open models on Meta’s own figures.
On StrongREJECT, a test of jailbreak attempts where lower is better, Muse Spark Open scored 0.5 while GLM 5.3 and Kimi K3 scored above 60. The same table shows the cost of that caution. On the FORTRESS test, Muse Spark Open refused harmless requests more often than any other model listed, with an over-refusal score of 17.4 against 2.7 for Kimi K3. If you plan a customer-facing assistant, test how often it says no to ordinary questions.
Two Mac Studios means about a terabyte of memory
Meta’s headline setup is two Mac Studios with Apple’s M5 Ultra chip, each with up to 512 GB of memory, linked through Apple’s MLX software. You cannot buy that setup yet: VentureBeat reported that the 512 GB version is expected in late October, and Apple has not published its price.
On servers, the model card asks for at least 700 GB of GPU memory in the compressed 4-bit format and 2 TB at full precision. Hugging Face ran a smaller 3-bit community version on one older M3 Ultra Mac with 512 GB. It produced about 21 tokens per second. That is enough for one person to test with, not for a team sharing one machine.
If you do not own that hardware, hosted versions are already live:
- Weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Hugging Face asks for your legal name, date of birth and organisation before you get access through the Muse Spark Open model card.
- Hosted APIs from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Together AI charges $1.25 per million input tokens and $4.25 per million output tokens, VentureBeat reported.
- Languages: more than 100 in training, but Meta does not list them.
The Muse licence is free, with a $20 million clause for API sellers
The Muse Model License lets you use, change, fine-tune and redistribute the model without paying Meta. Three conditions stand out:
- Products built on the model must show “Built with Muse Spark Open” somewhere visible, such as the documentation or a settings menu.
- A company that sells access to AI models as a service, and earned more than $20 million in any 12 months, needs a separate agreement with Meta first. Internal use and building the model into your own product are exempt.
- A model trained on Muse Spark Open’s answers counts as a derivative, so a smaller model you distil from it falls under the same licence.
The licence also has no clause that shuts out European companies. Meta’s Llama 4 licence withheld its rights for multimodal models from companies with their main office in the EU. For Muse Spark Open, Meta Platforms Ireland is simply the licensor for European users.
Muse Spark Open lets European teams avoid Meta’s hosted API and its missing EU region
For the organisations DataNorth works with, the deciding question is usually where the data goes, not which model tops a leaderboard. Meta has named no EU data region for its hosted Muse API. Muse Spark Open runs on your own servers or at a provider you choose, so client files and personal data never pass through Meta. For organisations whose policy rules out Chinese models, it is now the strongest American open model on Meta’s figures, with a wide lead over Nvidia’s Nemotron 3 Ultra.
Start with a hosted endpoint, not with hardware. Give Muse Spark Open two or three of your own agent tasks that call real tools. Include Dutch prompts if your users work in Dutch, because Meta does not say how well the model handles it. Check where each provider runs the model before you send personal data; the Dutch company Nebius is one of the launch hosts. Buy hardware only when data may not leave your network and you can budget for about a terabyte of fast memory. If coding is the main job and your policy allows Chinese models, GLM 5.3 and Kimi K3 still deliver more on Meta’s own table.
For more information, visit the official announcement of Muse Spark Open on the Meta AI Research blog.
