Meta Muse Spark Open, a mixture-of-experts Transformer released on October 15 2026 by Meta Superintelligence Labs, became available today as a fully open-weights model with 816 billion total parameters and 42 billion active parameters per token. The release introduces a 1 million token context window at inference and multimodal inputs spanning text, image, video and audio, with text-only output, positioning the model for long-horizon coding and agentic workflows alongside the co-trained Muse Code stack.
Meta Muse Spark Open Architecture And Modality Coverage
Meta Muse Spark Open ships with a vocabulary of 202,048 tokens spanning more than 100 languages, paired with a multimodal encoder pipeline that accepts text, image, video and audio inputs while constraining outputs to text. The mixture-of-experts backbone activates approximately 42 billion of the 816 billion total parameters on every token, an efficiency ratio Meta cites as central to the model’s throughput at long context lengths. Meta describes Meta Muse Spark Open as “the most capable US open-weights model to date,” a claim tied to the combination of compressed MoE routing and a hybrid global and local attention pattern that reduces inter-GPU traffic and KV cache memory pressure during extended agent sessions. The architecture is co-trained with the companion Muse Code stack, which feeds supervised signal from coding and tool-use trajectories directly into the base model rather than applying it as a separate post-training phase.
Meta Muse Spark Open Weights, License And Hardware Requirements
The weights for Meta Muse Spark Open are distributed under the Muse Model License, a commercial terms document that permits production deployment but requires a “Built with Muse Spark Open” attribution notice in derivative user interfaces. A separate written agreement with Meta becomes mandatory for any organization offering the model as a managed service and reporting more than $20 million in annual revenue. The repository ships BF16 and NVFP4 checkpoints, and the two formats determine very different memory footprints for self-hosting. NVFP4 deployment requires at least 700 GB of GPU VRAM, while BF16 deployment raises that threshold to at least 2 TB of GPU VRAM, numbers that push on-premise hosting toward multi-GPU server nodes or specialized inference appliances. For developers without that hardware, Meta confirms that two Apple Mac Studio M5 Ultra workstations configured with 512 GB of unified memory each can run Meta Muse Spark Open locally through the MLX framework, marking one of the largest local-runtime demonstrations for an open-weights model of this scale.
Meta Muse Spark Open Inference Stack And Day 0 Hosted Access
Open-source inference frameworks including SGLang, vLLM, Hugging Face Transformers, TensorRT-LLM, Unsloth and Llama.cpp received reference support at launch, with Meta publishing conversion scripts for the NVFP4 path alongside the BF16 checkpoints. For teams preferring hosted endpoints, day-zero access is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, a roster that covers major North American and European inference providers and shortens the path from download to served API. Meta’s recommended sampling parameters for Meta Muse Spark Open set Temperature at 1.0 and Top_p at 1.0 with Top_k disabled, while steering users to set Reasoning Effort to high or xhigh on complex coding, planning and tool-use tasks. The compressed MoE block design carries through to serving, where the combination of reduced all-to-all bandwidth and smaller per-layer KV footprints is intended to keep per-token latency flat as prompts approach the full 1 million token window.
Meta Superintelligence Labs trained the 816B-parameter Meta Muse Spark Open alongside Muse Code, an AI coding assistant positioned against Anthropic Claude Code and OpenAI Codex, and the company framed the November 2026 rollout as a US counterweight to Chinese open-weight leaders in the Artificial Analysis snapshot. By raw parameter count, Meta Muse Spark Open is the largest open-weight US model released to date, though Xiaomi’s MiMo-V2.6-Pro still tops the Artificial Analysis Index at 46.3, ahead of the Meta release’s score. Day-0 hosted endpoints span Baseten running NVFP4 on Blackwell silicon, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, giving developers multiple paths to wire the weights into production. The Muse Model License obliges any derivative product UI to carry a “Built with Muse Spark Open” notice, and MaaS operators crossing $20M in trailing 12-month gross revenue must sign a separate written agreement with Meta before exposing the model as a paid service. Meta’s published model card also states that Meta Muse Spark Open received additional safety post-training beyond the baseline Muse Spark release. Recommended inference settings are Temperature 1.0, Top_p 1.0, Top_k disabled, with Reasoning Effort set to high or xhigh for complex coding workloads.
Meta Muse Spark Open Release Context And Developer Backlash Response
The October 15 publication of Meta Muse Spark Open arrives two days after the October 13 launch of the closed Meta Muse Wonder developer tier and one day after coverage of the Muse Wonder developer backlash, in which outside developers criticized licensing terms attached to the proprietary model. By contrast, Meta Muse Spark Open is positioned as the open-weights counterpart, with permissive commercial rights for most deployment scenarios and visible reference deployments across mainstream inference frameworks. Internal framing frames the dual-track strategy as a way to capture both enterprise managed-service demand through Muse Wonder and community-driven adoption through Meta Muse Spark Open, which Meta Superintelligence Labs says is “the most capable US open-weights model to date” across published benchmarks at release.
Source: Hugging Face (https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B).