--- title: "Meta’s Muse Spark Open Lands With Fine Print and a China Gap" date: 2026-10-15 author: "Barry Elad" featured_image: "https://sqmagazine.co.uk/wp-content/uploads/2026/10/muse-spark-open-launched-with-apple.jpg" categories: - name: "Artificial Intelligence" url: "/artificial-intelligence.md" tags: - name: "News" url: "/tag/news.md" --- # Meta’s Muse Spark Open Lands With Fine Print and a China Gap Meta released the weights for Muse Spark Open on October 15, 2026, calling it the “**most capable US open-weight model to date.**” Meta’s own benchmarks still place it behind the top Chinese open-weight models. ## Key Points - Muse Spark Open has 816 billion parameters, with 42 billion active per request and a one million-token context window. - Meta says two linked Mac Studios running Apple’s M5 Ultra chip can run the full model on a desktop. - Xiaomi’s MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index, while Meta cites 39 for its own model. - AI service providers earning over $20 million in 12 months from the model must notify Meta and sign a separate deal. The model reads text, images and video natively, and Meta built it for coding and long-running agent work. In a [blog post](https://research.meta.ai/blog/muse-spark-open), Meta said it shipped bug fixes and new features to Ice Cubes, an open-source Mastodon client, entirely on-device. It also drove **DaVinci Resolve**, reviewing footage frame by frame to find and crop the strongest moments. **Apple’s MLX framework** carries the desktop claim. Each M5 Ultra Mac Studio lets its GPU use up to **512 GB** of unified memory, so two machines hold the whole model. According to Unite.AI, it needs at least 700 gigabytes of graphics memory, depending on format. That two-Mac setup costs about as much as a compact car. Everyone else can pull the weights from [Hugging Face](https://huggingface.co/collections/meta-models/muse-spark-open), **Amazon SageMaker,** [Google Cloud](https://sqmagazine.co.uk/google-cloud-platform-statistics/) or Microsoft Foundry. API access runs through Together AI, Fireworks, Groq and OpenRouter. > Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple. https://t.co/U2wJNrKBOg > > — Mark Zuckerberg (@finkd) [October 15, 2026](https://x.com/finkd/status/2110778025629347999?ref_src=twsrc%5Etfw) ## Apple Partnership is Surprising The Apple tie-up is the surprise, given years of privacy fights between the two firms. Meta AI chief Alexandr Wang told Axios the companies have worked together for several months. He said he’s “**excited for mommy and daddy to be playing nice again,**” and credited former Apple executive Daniel Gross with brokering it. The thaw follows Meta’s earlier [hiring of Apple AI chief Ke Yang](https://sqmagazine.co.uk/meta-hires-apple-ai-chief-ke-yang/). The release also reverses Meta’s spring move, when Muse Spark launched without open weights and broke from the Llama playbook. ## Meta’s own charts put Chinese labs ahead Nearly every benchmark so far comes from Meta. They show **Muse Spark Open** beating [Nvidia’s Nemotron 3 Ultra](https://sqmagazine.co.uk/nvidia-nemotron-3-5-super-vl/) and Thinking Machines’ Inkling, often by more than 20 points. | Benchmark | Muse Spark Open | Best Chinese model | |---|---|---| | **Artificial Analysis Intelligence Index** | 39 | 46 (MiMo-V2.6-Pro) | | **Terminal-Bench 2.1 (coding)** | 76.4 | 85.0 (Kimi K3) | | **GPQA Diamond (science)** | 89.0 | 93.5 (Qwen3.8, Kimi K3) | | **SWE-Atlas-QnA (software)** | 58 | 66 (Kimi K3) | Meta’s model does lead on agent and tool tasks, scoring **89.2** on MCP-Atlas and topping JobBench. Moonshot AI, which keeps shipping faster Kimi coding models, holds the coding edge with Kimi K3. Artificial Analysis hasn’t listed **Muse Spark Open** yet, so the 39 remains unconfirmed. Its open-weight ranking currently has Inkling as the top US entry, ninth with 25 points. If Meta’s score holds, Muse Spark Open would tie DeepSeek V4.1 Flash for fifth. Mistral Large 4 scored 38 in the same data, which leaves Meta and Mistral neck and neck for best open model outside China. ## The license draws a line at $20 million Meta ships the model under its custom Muse Model License, which goes further than Apache 2.0 or MIT. The license permits: - **Free use, copying and distribution** - **Fine-tuning, derivative models and distillation** - **Commercial products built on the model** - **Outputs, which Meta claims no rights over** The revenue cap (about €18 million) targets large hosts, and internal use, research and embedded products stay exempt. Every product using the model or a derivative must display “**Built with Muse Spark Open**” in its interface, documentation or settings. The usage policy also bans weapons work, malware and fraud, and the model isn’t meant for anyone under 18. ## What’s Next? Teams already testing it should check whether their product shows that notice. Those planning a hosted service should model when revenue could cross the cap. The real test now sits with Artificial Analysis. Its listing, plus Meta’s full evaluation report, will show whether that 39 holds and a US lab finally reaches the open-weight top five. Definition of Model Weights. Link to full glossary entry follows the description.**Model Weights**Model weights are the learned parameters a trained AI model uses to turn an input into an output, and the artifact a lab chooses to publish or withhold. [Read more about Model Weights](https://sqmagazine.co.uk/glossary/model-weights/) Definition of Context Window. Link to full glossary entry follows the description.**Context Window**A context window is all the text an AI model can reference when generating a response, measured in tokens and shared with the model's own output. [Read more about Context Window](https://sqmagazine.co.uk/glossary/context-window/)