Muse Spark Open 816B: Meta's Open MoE Model

Meta Superintelligence Labs has released Muse Spark Open, the open-weight version of its Muse Spark model. This 816-parameter Mixture-of-Experts architecture activates 42 billion parameters per token and maintains a 1-million-token context window. It accepts text, image, video, and audio input and returns text output. Meta positions it as “the most capable US open-weight model to date,” a qualifier that reflects its strength in certain domains while acknowledging gaps in others.

What changed

Meta has published Muse Spark Open weights across multiple platforms, including Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Two distribution formats are available: BF16 (approximately 1.6 TB) for full precision and NVFP4 quantization (approximately 530 GB). The open version underwent additional safety training before release.

The architecture spans 62 layers with a hidden dimension of 8,192. Each Mixture-of-Experts layer contains 256 experts, with top-k routing activating 8 per token. A separate 1.8-billion-parameter Perception Encoder processes visual input. Grouped-query attention employs 128 query heads and 16 key-value heads. The vocabulary contains 202,048 tokens and was trained on data covering more than 100 languages. Users can configure reasoning depth through a parameter with four settings: low, medium, high, and xhigh (the default).

Why it matters

For teams developing agentic workflows, Muse Spark Open reduces deployment complexity. The model runs on two Apple M5 Ultra Mac Studios connected via Thunderbolt 5 with Remote Direct Memory Access. Each machine provides 512 GB of unified memory. MLX tensor parallelism distributes each layer across both machines. This represents the first model in its performance class advertised to run on consumer Mac hardware at scale, making advanced agentic capabilities available to developers with existing Apple Silicon workstations.

Data center operators have multiple deployment paths. NVFP4 quantization runs on 4 Blackwell Ultra GPUs or 8 H200 accelerators. Workstation deployments using Unsloth’s 2-bit GGUF quantization require approximately 300 GB of combined memory. Seven hosted APIs provide immediate access: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. OpenRouter pricing sits at $1.25 per million input tokens and $4.25 per million output tokens.

Inference support launches across multiple engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth. Meta included custom Helion GPU kernels in SGLang, reporting kernel-level speedups ranging from 1.03x to 3.95x on NVIDIA and AMD GPUs. Fine-tuning is supported through TorchTitan, Hugging Face TRL, and NVIDIA NeMo AutoModel.

For developers comparing open-weight alternatives to frontier models, smaller options exist. However, Muse Spark Open’s 1-million-token context and Mixture-of-Experts architecture target specific use cases for long-horizon reasoning and extended workflows.

What to test

Benchmark performance warrants careful evaluation before production adoption. Meta’s published comparison table shows competitive performance on tool use and knowledge tasks but reveals coding performance gaps.

On MCP-Atlas (tool use), Muse Spark Open scores 89.2, exceeding Qwen 3.8 (84.5) and GLM 5.3 (84.2). On JobBench (job reasoning), it reaches 61.8, slightly ahead of GLM 5.3 (61.4). On AA Omniscience, it achieves the highest score at 25. These results indicate strong agentic and reasoning capabilities.

Coding performance presents the main weakness. On Terminal-Bench 2.1 with the Terminus 2 harness, Muse Spark Open scores 76.4. Kimi K3 reaches 85.0, GLM 5.3 scores 83.9, and Qwen 3.8 scores 82.0. Its Artificial Analysis Intelligence Index stands at 39, behind GLM 5.3 (45), Kimi K3 (44), and Qwen 3.8 (40). Some rival scores are self-reported, so independent verification strengthens production deployment decisions.

Practical testing should measure five dimensions before committing to adoption. First, measure throughput on target hardware, comparing MLX performance on Mac Studios against NVFP4 on GPUs. Second, evaluate latency for multi-turn agentic workflows involving tool calling. Third, assess accuracy on domain-specific coding tasks relevant to your workload. Fourth, profile memory footprint across different quantization levels. Fifth, compare cost-per-token across hosted APIs and self-hosted inference options.

The licensing terms significantly affect commercial deployment. The Muse Model License permits use, fine-tuning, distillation, and redistribution. Two key conditions apply: products must display “Built with Muse Spark Open,” and Model-as-a-Service providers with more than $20 million in 12-month revenue require a separate commercial agreement with Meta. Internal use, research, and powering your own product features remain exempt from the revenue threshold.

Common questions

Is Muse Spark Open truly open-weight?

Yes. Weights are publicly available on Hugging Face in BF16 and NVFP4 formats. The model ships under the Muse Model License, which permits downloading, fine-tuning, distillation, and redistribution. Most developers can use it without additional licensing. Only Model-as-a-Service providers exceeding $20 million in annual revenue require a separate agreement with Meta.

How do I run Muse Spark Open on my Mac?

Use the Apple MLX framework with two M5 Ultra Mac Studios, each with 512 GB unified memory, connected via Thunderbolt 5. MLX tensor parallelism distributes computation across both machines. Alternatively, use Unsloth’s 2-bit GGUF quantization fitting approximately 300 GB of combined memory on other hardware, or access hosted APIs from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

How does Muse Spark Open compare to other open models?

On tool use and reasoning benchmarks, it leads most US open models. On Terminal-Bench 2.1 coding, Chinese open models (Kimi K3, GLM 5.3, Qwen 3.8) demonstrate superior performance. Meta’s claim of “most capable US open-weight model” reflects this regional distinction. Benchmark performance varies by task, so evaluation against your specific workload is essential before adoption.

The conclusion

Muse Spark Open 816B addresses a genuine need: a large, multimodal open-weight model that runs on accessible hardware. Its 1-million-token context and Mixture-of-Experts design serve long-horizon agentic workflows. Deployment flexibility spanning Mac Studios, data center GPUs, and hosted APIs lowers barriers for teams at different scales.

The performance trade-off is distinct. This model excels at tool use, reasoning, and knowledge tasks. Coding performance trails open models from other regions. If your workload prioritizes agentic orchestration over code generation, Muse Spark Open merits testing. If code quality is central to your application, evaluate both agentic workflow patterns and model selection in conjunction.

Monitor community activity: fine-tunes on Hugging Face optimized for specific domains, independent benchmark verification, latency and cost comparisons across hosted platforms, and whether smaller alternatives serve your use case better. The licensing terms are favorable for most teams, so the key question becomes technical fit rather than cost.

Related reading

• Meta Muse Wonder frontier model launches today
• Nemotron 3.5 Super VL 120B: Open Multimodal Agent Model
• RTX Spark PC AI launches Oct 16 with local agents