Muse Spark Open arrives as the open-weight release of Meta Superintelligence Labs’ flagship model. It is an 816B-parameter Mixture-of-Experts with 42B active parameters per token. The model accepts text, images, video, and audio, and returns text. Its context window reaches one million tokens. The Meta team calls it the most capable US open-weight model to date. For developers, deployment is the headline: the NVFP4 build runs on two Apple M5 Ultra Mac Studios. You get scale, multimodality, and long context without closed weights.
What is Muse Spark Open and why does it matter?
Muse Spark Open is a natively multimodal MoE model from Meta for coding and agentic work. It processes text, images, video, and audio, and answers in text. A one million token context unlocks long tasks and sustained action chains. The Meta team claims top capability among US open-weight models.
Weights ship in BF16 and NVFP4 on Hugging Face. The open version received extra safety post-training. Meta states that this improves robustness without system-level guardrails. That reduces heavy platform filters in practice.
The quick spec lists 816B total and 42B active parameters. The context is 1M tokens. Strongest areas include tool use and knowledge reliability in Meta’s tables. Coding performance sits mid-pack among large open models.
For developers, the practical leap is running on two M5 Ultra Mac Studios. The NVFP4 build with MLX splits layers across both machines. That pairs desktop hardware access with heavyweight model performance. Ready to dive into the architecture?
“Muse Spark Open is the most capable US open-weight model,” the Meta team states.
How does the architecture work and what does MoE imply?
The stack has 62 layers, a hidden size of 8192, and MoE with 256 experts. Top‑k routing activates eight experts per token, delivering 42B active parameters. Attention is grouped‑query with 128 query heads and 16 KV heads. A separate 1.8B‑parameter Perception Encoder handles visual input.
The vocabulary contains 202,048 tokens. Training data covers more than 100 languages. This supports mixed domains and multimodality without separate models. MoE provides total scale without linear inference cost growth.
Reasoning depth is controlled by the reasoning_effort setting. Available values are low, medium, high, and xhigh. The default is xhigh. That favors quality when latency budgets allow.
Meta positions the model for coding and long‑horizon agent chains. The mix of GQA, long context, and MoE serves that goal. The dedicated perception encoder takes over image inputs. Wondering how this plays out in benchmarks?
The “reasoning_effort” options: low, medium, high, or xhigh (default is xhigh).
How does it perform on benchmarks?
It is strongest on tool use and knowledge reliability. On MCP‑Atlas it scores 89.2, ahead of Qwen3.8 (84.5) and GLM 5.3 (84.2). On JobBench it reaches 61.8, edging GLM 5.3 at 61.4. On HLE‑Diamond without tools it posts 26.1 versus 22.2 for Kimi K3. On AA Omniscience it records 25, the top result in Meta’s table.
With tools, ZeroBench shows 50.0 versus 41.0 for Kimi K3. Coding is the softer spot. On Terminal‑Bench 2.1 with the Terminus 2 harness it scores 76.4. Kimi K3 hits 85.0, GLM 5.3 gets 83.9, and Qwen3.8 scores 82.0. Its Artificial Analysis Intelligence Index is 39. That trails GLM 5.3 (45), Kimi K3 (44), and Qwen3.8 (40).
The takeaway is direct: “most capable US” holds. Yet open Chinese models often lead on aggregate measures. Some rival scores are self‑reported. The evaluation methodology lists each source. Against that backdrop, deployment simplicity matters even more.
MCP‑Atlas: 89.2. JobBench: 61.8. ZeroBench with tools: 50.0. Terminal‑Bench 2.1: 76.4. AAI Index: 39.
What hardware is required, and how do you run it locally?
BF16 weights are about 1.6 TB, while NVFP4 lands near 530 GB. For NVFP4 the model card recommends at least 700 GB of VRAM. For BF16 it suggests about 2 TB. That frames both data center and workstation plans.
In data centers, NVFP4 runs on four Blackwell Ultra GPUs or eight H200s. On desktops, the MLX build runs on two M5 Ultra Mac Studios with 512 GB each. The machines connect over Thunderbolt 5 with RDMA. MLX applies tensor parallelism, splitting every layer across both nodes.
For workstations, Unsloth Dynamic GGUFs go down to 2‑bit. They fit in roughly 300 GB of combined memory. The headline states simply: “runs on two Mac Studios.” That lowers the entry barrier for long context and multimodality. Want to see where else you can run it today?
Runs on: 2 M5 Ultra Mac Studios (512 GB each) via MLX; or 4 Blackwell Ultra / 8 H200s in data centers.
Where can you run it, how to tune it, and what does the license allow?
Day‑one engines include SGLang, vLLM, TensorRT‑LLM, Transformers, MLX, llama.cpp, and Unsloth. Hosted APIs are live on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. OpenRouter lists $1.25 per 1M input tokens and $4.25 per 1M output tokens. That enables quick trials without running your own stack.
Two draft modules accelerate decoding: DFlash V2 for SGLang and DSpark for vLLM. Meta reports 3.7× single‑request throughput with DFlash V2 on AMD MI355X. Fine‑tuning works with TorchTitan, Hugging Face TRL, and NVIDIA NeMo AutoModel. That covers inference and domain adaptation paths.
The Muse Model License permits use, fine‑tuning, distillation, and redistribution. Two conditions matter most. First, products must display “Built with Muse Spark Open.” Second, Model‑as‑a‑Service providers above $20M in 12‑month revenue need a separate agreement with Meta. Internal use, research, and powering your own features are exempt.
Head‑to‑head comparisons are clear. Total params: Muse Spark Open at 816B, Kimi K3 at 2.8T, GLM 5.3 at 753B, Qwen3.8‑2.4T‑A95B at 2.4T. Active params: 42B versus 104B, about 40B, and 95B respectively. Context: 1M for Muse, Kimi, and GLM; Qwen3.8 has 262,144 native, up to 1,010,000. Input: Muse supports text, image, video, audio; Kimi supports text, image, video; GLM and Qwen focus on text. Released weights: BF16 and NVFP4 for Muse; MXFP4 (QAT) for Kimi; BF16 and FP8 for GLM; BF16 for Qwen. AAI Index: 39 for Muse, 44 for Kimi, 45 for GLM, 40 for Qwen. Terminal‑Bench 2.1: 76.4 for Muse, 85.0 for Kimi, 83.9 for GLM, 82.0 for Qwen. MCP‑Atlas: 89.2 for Muse, 82.3 for Kimi, 84.2 for GLM, 84.5 for Qwen. Toolathlon Verified (vendor): 72.8 for Muse, 76.5 for Kimi, 73.0 for GLM, 72.5 for Qwen. HLE with tools (vendor): 61.0 for Muse, 56.0 for Kimi, 62.5 for GLM, 56.2 for Qwen. DeepSWE 1.1 (vendor, own harness): 60 for Muse, 67.5 for Kimi, 66.9 for GLM, 56.6 for Qwen. The table lists a desktop setup only for Muse: 2× M5 Ultra Mac Studio. Ready to try your own cases?
“Built with Muse Spark Open” is mandatory labeling for shipped products.
Based on Meta Superintelligence Labs.
