Meta Releases Muse Spark Open 816B-Parameter Coding Model
The 816-billion-parameter mixture-of-experts model brings multimodal reasoning, a 1-million-token context window, and local execution on paired Mac Studio hardware.
In short
- Meta released Muse Spark Open, an 816-billion-parameter open-weights model activating 42 billion parameters per token with a 1-million-token context window.
- The model processes text, image, document, audio, and video inputs natively, and generates explanations before taking actions in complex coding workflows.
- Local desktop deployment requires 1TB of unified memory, achievable with two Apple M5 Ultra Mac Studios using Apple's MLX framework.
- On agentic evaluations like JobBench (61.8) and MCP-Atlas (89.2), the model leads open-weight alternatives, but trails top Chinese models on broader intelligence indices.
- Available through Hugging Face, AWS SageMaker JumpStart, and third-party APIs under the Muse Model License, which includes commercial service restrictions.
Meta has launched Muse Spark Open, an 816-billion-parameter open-weights model designed specifically for agentic software engineering and multimodal reasoning. Meta announced Muse Spark Open from Meta Superintelligence Labs on October 15, 2026.
Meta describes Muse Spark Open as the most capable U.S. open-weight model to date. Built to tackle sustained, multistep programming tasks rather than simple code autocompletion, the model incorporates a 1-million-token context window that enables it to ingest entire repositories, external libraries, and historical logs simultaneously. Alongside cloud deployments, Meta demonstrated the model running locally on paired Mac Studio desktop workstations.
Architecture and Multimodal Perception
Muse Spark Open utilizes a sparse mixture-of-experts (MoE) transformer architecture. According to the official Hugging Face model card, the system routes each token to eight of 256 available experts, activating 42 billion of its 816 billion total parameters per forward pass. The network spans 62 layers with 128 query attention heads, 16 key-value heads using grouped-query attention, an 8,192 hidden dimension, and a vocabulary size of 202,048 tokens.
Unlike traditional text-only coding engines, Muse Spark Open is natively multimodal. A 1.8-billion-parameter perception encoder with 50 layers processes interleaved text, images, video frames, and audio tracks in a single unified pass. Meta notes that this unified perception allows the system to inspect UI screenshots, parse architectural diagrams, inspect error logs, and diagnose software defects directly from narrated screen recordings.
During execution, the model uses goal conditioning and context compaction to maintain trajectory across extended programming sessions. It outputs deliberate reasoning and intermediate commentary before executing tool calls or applying code patches, allowing engineers to verify its planning process.
Benchmark Performance Against Open-Weight Rivals
Meta positions Muse Spark Open as an answer to open-weight models from Chinese AI laboratories, which have gained significant developer traction. According to figures compiled by VentureBeat, Chinese models represented 41% of Hugging Face downloads in the year leading to March 2026 and accounted for the majority of open-model token traffic on OpenRouter in September 2026.
On Meta's published benchmarks, evaluated at maximum reasoning effort, Muse Spark Open outperformed international peers on key agentic tasks. It achieved 61.8 on JobBench, edging past GLM 5.3 (61.4), Kimi K3 (54.3), and Qwen3.8-2.4T-A95B (52.7). On the MCP-Atlas tool benchmark, Muse Spark Open scored 89.2, ahead of GLM 5.3's 84.2 and Qwen3.8's 84.5. On knowledge evaluations without tools, it reached 26.1 on HLE-Diamond, compared to 22.2 for Kimi K3 and 17.3 for Qwen3.8.
However, broader multi-discipline evaluations tell a more nuanced story. On the Artificial Analysis Intelligence Index, Muse Spark Open scored 39, trailing GLM 5.3 (45), Kimi K3 (44), and Qwen3.8 (40), while remaining ahead of domestic open competitors such as Inkling (26) and Nemotron 3 Ultra (23). In coding evaluations like DeepSWE 1.1, Muse Spark Open recorded a score of 60, trailing Kimi K3 (68) and GLM 5.3 (67). On Terminal-Bench 2.1, its 76.4 score was surpassed by both Kimi K3 (85.0) and GLM 5.3 (83.9).
Hardware Demands and Desktop Execution
Serving an 816-billion-parameter model imposes severe memory requirements. While sparse activation limits the compute expended per token to that of a 42-billion-parameter model, holding the complete network in memory requires substantial capacity. According to Meta's specifications, the full BF16 checkpoint requires at least 2TB of accelerator memory, while the 4-bit NVFP4 quantized checkpoint requires at least 700GB.
In enterprise environments, the NVIDIA Technical Blog reports that Muse Spark Open was trained on NVIDIA GB300 NVL72 Blackwell Ultra systems. On four NVIDIA Blackwell GPUs using speculative decoding, the model delivers over 50,000 tokens per second per GPU and up to 600 tokens per second per user, benefiting from 130 TB/s NVLink inter-GPU bandwidth.
For local desktop environments, Meta collaborated with Apple to optimize the model for Apple Silicon using the open-source MLX framework. As covered by MacRumors and 9to5Mac, two Mac Studio units equipped with M5 Ultra processors and 512GB of unified memory can be linked to act as a single 1TB unified memory pool. Demonstrations conducted by Meta showed the local cluster modifying the open-source Ice Cubes Mastodon iOS client and automating DaVinci Resolve to edit video footage.
Running this desktop setup represents a substantial investment. Apple's 512GB M5 Ultra Mac Studio is slated for release in late October 2026, with official pricing unannounced. Engadget notes that an alternative four-machine cluster of 256GB M5 Ultra units costs roughly $38,000, whereas VentureBeat reported the $43,196 calculation before taxes and accessories.
Licensing, Cloud Availability, and Pricing
Meta is distributing Muse Spark Open under the custom Muse Model License. The license permits local deployment, fine-tuning, and redistribution, but includes strict commercial protections. Organizations that generate more than $20 million in aggregate gross revenue over any consecutive 12-month period must notify Meta and obtain a separate written agreement before offering the model or its derivatives as a commercial model-as-a-service. Internal enterprise deployments and integrations into software applications are exempt from this threshold. Products incorporating the model must display the attribution "Built with Muse Spark Open."
Downloading weights from Hugging Face requires completing a gated access form that mandates users submit their full legal name, date of birth, and complete corporate identifiers without acronyms.
For managed infrastructure, the model is available on Amazon SageMaker JumpStart, where the NVFP4 variant runs on ml.p5e.48xlarge instances and the BF16 variant runs on ml.p6-b300.48xlarge instances. Weights are also hosted on Google Cloud Model Garden and Microsoft Foundry Models. Serverless API hosting is available via Together AI at $1.25 per million input tokens ($0.15 for cached inputs) and $4.25 per million output tokens, alongside providers including Baseten, Fireworks AI, Groq, Modal, Nebius, and OpenRouter.
Questions people ask
What are the parameter counts and context length of Muse Spark Open?
Muse Spark Open features 816 billion total parameters in a mixture-of-experts configuration, with 42 billion active parameters per token, and supports a context window of up to 1 million tokens.
Can Muse Spark Open run on local consumer hardware?
No; paired 512GB M5 Ultra Mac Studios are high-end professional workstations costing tens of thousands of dollars, not standard consumer hardware, and local execution requires at least 1TB of memory. Meta optimized the model to run locally across two Apple M5 Ultra Mac Studios with 512GB of unified memory each using Apple's MLX framework.
What inputs and outputs does the model support?
Muse Spark Open natively accepts text, images, documents, audio, and video inputs via an integrated 1.8B perception encoder. It generates text outputs only.
What commercial restrictions exist under the Muse Model License?
Licensees generating over $20 million in gross revenue over 12 months require a separate agreement from Meta to provide the model as a paid model-as-a-service. Internal use and non-service product integrations are exempt.
Sources
- Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72NVIDIA Technical Blog, Oct 15, 2026, official source
- Meta’s Muse Spark Open now available in Amazon SageMaker JumpStart | Amazon Web ServicesAmazon Web Services, Oct 15, 2026, official source
- unsloth/Muse-Spark-Open-816B-A42BHugging Face, Sep 18, 2026, official source
- Meta unveils Muse Spark Open, the new top U.S. open weights model — and it runs on two paired Mac desktopsventurebeat.com, Oct 15, 2026
- Meta and Apple Bring Muse Spark Open AI Model to Mac StudioMacRumors, Oct 15, 2026
- Meta Built A Huge Open-Weight AI Model For Macs, But You'll Need 1TB Of RAM To Run ItEngadget, Oct 15, 2026
- Meta launches new Muse Spark Open model, optimized to run on Mac Studio9to5Mac, Oct 15, 2026
How this story was made: the newsroom picked it up from Google News, Techmeme and engadget.com, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (35 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 16, 2026 at 00:15 UTC
