Meta's Muse Spark Open Puts 816B MoE on Two Mac Studios

Meta Superintelligence Labs released Muse Spark Open on October 15: an 816B-parameter open-weight mixture-of-experts model with 42B active parameters per token and a 1M-token context window. The model accepts text, image, video, and audio and returns text. Meta describes it as the most capable US open-weight model to date, but the more actionable claim for engineers is deployment: the NVFP4 checkpoint runs on two M5 Ultra Mac Studios, and MarkTechPost reports hosted APIs already live on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

Architecture and modality

Meta's model card lists 62 layers with a hidden dimension of 8192. Each mixture-of-experts layer contains 256 experts, with top-k routing activating 8 per token. Attention uses grouped-query attention with 128 query heads and 16 KV heads. A separate 1.8B-parameter Perception Encoder handles visual input. The tokenizer has 202,048 tokens and training data spans more than 100 languages. Reasoning depth is controlled through reasoning_effort with low, medium, high, and xhigh settings; the default is xhigh. Weights ship in BF16 and NVFP4. Meta says the open version received extra safety post-training, making it more robust without system-level guardrails—a vendor-reported safety claim, not yet independently verified.

Benchmarks: tool use first, coding mid-pack

Meta's comparison table shows Muse Spark Open strongest in agentic tool use and knowledge reliability. It reports MCP-Atlas at 89.2, ahead of Qwen3.8-2.4T-A95B at 84.5 and GLM 5.3 at 84.2. JobBench lands at 61.8, narrowly above GLM 5.3's 61.4. On HLE-Diamond without tools, the model scores 26.1 versus 22.2 for Kimi K3. ZeroBench with tools reaches 50.0 versus 41.0 for Kimi K3. On AA Omniscience it scores 25, the top score in Meta's table. Coding is the weak spot: Terminal-Bench 2.1 with the Terminus 2 harness scores 76.4, behind Kimi K3 at 85.0, GLM 5.3 at 83.9, and Qwen3.8 at 82.0. Meta's comparison puts the Artificial Analysis Intelligence Index at 39, below GLM 5.3's 45, Kimi K3's 44, and Qwen3.8's 40. MarkTechPost notes the "most capable US" claim holds, but Chinese open models still lead overall.

Meta's model card is the source for these figures; some rival scores are self-reported, and the evaluation methodology lists each source. Treat all of them as vendor-reported.

Deployment, pricing, and license

BF16 weights are about 1.6 TB. The NVFP4 checkpoint quantizes only the experts and lands near 530 GB, though Meta's model card recommends at least 700 GB of VRAM for NVFP4 and 2 TB for BF16. Data center serving runs on four Blackwell Ultra GPUs or eight H200s. The desktop path is two M5 Ultra Mac Studios with 512 GB each, linked over Thunderbolt 5 with RDMA; MLX tensor parallelism splits every layer across both machines, and four 256 GB Mac Studios also work. Unsloth Dynamic GGUF builds down to 2-bit fit in about 300 GB of combined memory. Day-one serving engines include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp, and Unsloth. OpenRouter lists $1.25 per 1M input tokens and $4.25 per 1M output tokens. Two draft models aim to speed decoding: DFlash V2 for SGLang and DSpark for vLLM. Meta reports 3.7x single-request throughput with DFlash V2 on AMD MI355X. Fine-tuning is supported through TorchTitan, Hugging Face TRL, and NVIDIA NeMo AutoModel.

The Muse Model License permits use, fine-tuning, distillation, and redistribution. Products must display "Built with Muse Spark Open." Model-as-a-Service providers above $20M in 12-month revenue need a separate agreement with Meta; internal use, research, and powering your own product features are exempt from the revenue clause.

AI Mastery analysis

The release changes local-inference math more than benchmark rankings. A 530 GB NVFP4 checkpoint fits in the 1 TB unified memory of two 512 GB M5 Ultra Mac Studios, but Meta's 700 GB VRAM recommendation is the planning number; the gap covers activations and KV cache before a 1M-token sequence is ever filled. That makes the model's 1M context as much a memory and pricing problem as a capability race. The license carries a similar boundary: open-weight is not unrestricted Model-as-a-Service, and the $20M trailing 12-month revenue clause inserts a commercial renegotiation exactly at scale. For model selection, the benchmark pattern is the strongest signal: Meta tuned for agentic tool orchestration and multi-step work, not competitive coding. If Terminal-Bench is the workload, several Chinese open models lead; if MCP-Atlas and JobBench matter more, Muse Spark Open is the more relevant open-weight candidate.

Muse Spark Open is less a benchmark sweep than a deployment statement. Frontier-scale open-weight multimodal MoE now fits on two desktop-class workstations, with commercial terms that permit most product use but impose a threshold for the largest providers. That continues the 2026 shift from raw capability scores toward memory budgets, serving engines, and license terms as the real production selection criteria.

Sources

• Primary: meta-models/Muse-Spark-Open-816B-A42B · Hugging Face
• Reported by: MarkTechPost: Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context