Meta Releases Muse Spark Open: 816B Open-Weight MoE Runs on Two Mac Studios

Meta Superintelligence Labs has released Muse Spark Open, an 816B-parameter open-weight Mixture-of-Experts model that activates 42B parameters per token and holds a 1M-token context window. The Meta team claims it is the most capable US open-weight model to date. For developers, the release carries a second headline: the NVFP4 build runs on 2 Apple M5 Ultra Mac Studios.

The model takes text, image, video and audio input and returns text. Meta built it for coding and long-horizon agentic work, and shipped the weights in BF16 and NVFP4 on Hugging Face. The open version received extra safety post-training. The Meta team states that this makes it more robust without system-level guardrails.

The release matters beyond Meta. Open-weight competition has shifted to the front of the industry agenda, and Chinese labs — Moonshot AI, Z.ai and Alibaba — currently publish the strongest-performing downloadable models. Muse Spark Open gives US developers a locally deployable contender, with hardware requirements low enough for a desktop setup rather than a data center rack.

What is the architecture?

The model card lists 62 layers and a hidden dimension of 8192. Each MoE layer holds 256 experts, and top-k routing activates 8 per token. Attention is grouped-query, with 128 query heads and 16 KV heads.

A separate 1.8B-parameter Perception Encoder handles visual input. The vocabulary spans 202,048 tokens, and training data covers more than 100 languages.

Developers control reasoning depth through a reasoning_effort setting: low, medium, high or xhigh. The default is xhigh.

How does it perform on benchmarks?

Meta compared Muse Spark Open with Kimi K3, GLM 5.3, Qwen3.8-2.4T-A95B and others at maximum reasoning effort. Its strongest results sit in tool use and knowledge reliability:

• MCP-Atlas: 89.2, ahead of Qwen3.8 (84.5) and GLM 5.3 (84.2) — the top score in Meta's comparison table.
• JobBench: 61.8, narrowly ahead of GLM 5.3 (61.4).
• HLE-Diamond, no tools: 26.1, versus 22.2 for Kimi K3.
• AA Omniscience: 25, the top score in Meta's table.
• ZeroBench with tools: 50.0, versus 41.0 for Kimi K3.

Coding is the weak spot. 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 Qwen3.8 scores 82.0.

The model's Artificial Analysis Intelligence Index is 39. That sits behind GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8 at 40. The claim of "most capable US" open-weight model holds; Chinese open models still lead overall.

Some rival scores are self-reported. Meta's published evaluation methodology lists the source for each number.

On other agentic benchmarks from the comparison table, Muse Spark Open scores 72.8 on Toolathlon Verified (Kimi K3: 76.5, GLM 5.3: 73.0), 61.0 on HLE with tools (GLM 5.3: 62.5) and 60 on DeepSWE 1.1 under its own harness (Kimi K3: 67.5, GLM 5.3: 66.9).

What hardware does it need?

Per Meta's developer guide, BF16 weights occupy about 1.6 TB. NVFP4 quantizes only the experts and lands near 530 GB. The model card recommends at least 700 GB of VRAM for NVFP4 and 2 TB for BF16.

• Data center: NVFP4 on 4 Blackwell Ultra GPUs or 8 H200s.
• Desktop: 2 M5 Ultra Mac Studios with 512 GB each, linked over Thunderbolt 5 with RDMA. MLX tensor parallelism splits every layer across both machines.
• Workstation: Unsloth Dynamic GGUFs down to 2-bit fit in about 300 GB of combined memory.

The desktop path is the differentiator. No rival in Meta's comparison table discloses a comparable consumer-hardware setup, and Kimi K3, GLM 5.3 and Qwen3.8 each list their desktop configurations as not disclosed.

Where can developers run it?

Day-one inference engines include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth. Hosted APIs are live on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

OpenRouter lists the model at $1.25 per 1M input tokens and $4.25 per 1M output tokens.

Two draft models speed up decoding: DFlash V2 for SGLang and DSpark for vLLM. Meta reports 3.7x single-request throughput with DFlash V2 on an AMD MI355X. Fine-tuning works with TorchTitan, Hugging Face TRL and NVIDIA NeMo AutoModel.

What does the Muse Model License allow?

The Muse Model License permits use, fine-tuning, distillation and redistribution. Two conditions matter most:

• 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 a company's own product features are exempt from the revenue clause.

The condition mirrors the commercial carve-out structure Meta used in earlier open-weight releases: free access for most builders, a negotiated deal for rivals reselling model access at scale.

How does it compare with Kimi K3, GLM 5.3 and Qwen3.8?

Feature	Muse Spark Open	Kimi K3	GLM 5.3	Qwen3.8-2.4T-A95B
Total params	816B	2.8T	753B	2.4T
Active params	42B	104B	~40B	95B
Context	1M	1,048,576	1,048,576	262,144 native, up to 1,010,000
Input	Text, image, video, audio	Text, image, video	Text	Text
Released weights	BF16, NVFP4	MXFP4 (QAT)	BF16, FP8	BF16
AA Intelligence Index	39	44	45	40
Terminal-Bench 2.1	76.4	85.0	83.9	82.0
MCP-Atlas	89.2	82.3	84.2	84.5

Muse Spark Open is the only model in the table with native audio input, and it activates fewer parameters per token than every rival except GLM 5.3. That efficiency gap explains the two-Mac-Studio deployment claim.

What happens next?

Muse Spark Open establishes a US contender in the open-weight race, but the benchmark spread — first on MCP-Atlas, last among large open models on Terminal-Bench 2.1 — defines its use case: agentic tool work over raw coding. Developers can test that claim directly, since the weights, research blog, developer guide and evaluation methodology are all public on day one.