Muse Spark Open: Meta's 816B open-weight model fits on two Mac Studios, with conditions
Meta released Muse Spark Open, an 816B mixture-of-experts model with a 1M-token context, under its own Muse Model License. Meta's card shows it leading some agentic tests and trailing Kimi K3 and GLM 5.3 on others. The numbers and terms are Meta's.
Facts checked October 16, 2026
On October 15 Meta released the weights of Muse Spark Open: 816B total parameters, 42B active per token, a 1M-token context, multimodal input. Meta says it runs on two Apple M5 Ultra Mac Studios through MLX. The license is the Muse Model License, permissive but not OSI open source. I read Meta's blog, the Hugging Face model card and the vLLM day-0 post. Meta release blog
What Meta released
Muse Spark Open is a sparse mixture-of-experts model: 816B total parameters, 42B active per token, 62 layers, 256 experts per layer with 8 routed per token, a 1M-token context at inference. Input is text, image, video and audio; output is text. The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry, in BF16 and NVFP4, with MXFP4 weights released by AMD. API access runs through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Model card
The two-Mac claim
Meta's headline for hobbyists is local inference: the company says Muse Spark Open is the first model in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with the latest MLX. Each M5 Ultra gives the integrated GPU up to 512 GB of unified memory, so two machines hold the model, and the 1.2 TB/s memory bandwidth feeds it. Meta's on-device demos used Muse Code powered by the model to ship real iOS changes to Ice Cubes, an open-source Mastodon client, and to drive DaVinci Resolve end to end, picking frames and cropping vertically.
These are Meta's own demos. In what I read, Meta published no tokens-per-second figure for the two-Mac setup, and no one outside Meta has replicated it yet. The hardware cost of two M5 Ultra Mac Studios is also left to the reader.
How 1M tokens became affordable
The vLLM team's day-0 post, written with Meta Superintelligence Labs, explains the architecture. Every one of the 62 layers is a mixture-of-experts layer. A learned pooler matrix halves each token's width from 8192 to 4096 before the experts and an unpooler restores it, which halves the parameters in the expert projections. Attention alternates on a cycle of four: three layers attend to a 2048-token sliding window, the fourth sees the whole context, and the cycle is anchored so the final layer is full attention. Across 62 layers that gives 46 narrow and 16 wide layers, and at 1M tokens the KV cache is just over a quarter of what all-full attention would need. That is what makes the context length cheap enough to serve.
vLLM also trained a speculative decoding draft model called DSpark, which proposes up to seven tokens per pass and works on CUDA and ROCm. Day-0 support ships as prebuilt vLLM images for NVIDIA and AMD; the code PR was still under review, with the nightly PyPI wheel expected about a week later. vLLM day-0 post
The benchmark table, read plainly
Meta's card compares Muse Spark Open at xhigh reasoning effort with Inkling, Nemotron 3 Ultra, Beam, Mistral Large 4, Qwen3.8-2.4T-A95B, GLM 5.3 and Kimi K3. Muse Spark Open posts the best number on MCP-Atlas (89.2, next 84.5) and the lowest, meaning best, StrongREJECT v2 score (0.5 against 62.9 to 86.4). It is close to the front on HLE with tools (61.0 against GLM 5.3's 62.5).
On other rows Meta's own table puts it behind. Terminal-Bench 2.1 with the Terminus 2 agent: 76.4 against Kimi K3's 85.0 and GLM 5.3's 83.9. DeepSWE 1.1: 60 against 68 and 67. GPQA Diamond: 89.0 against 93.5. GDPVal-AA v2: 1401 against 1653. The claim "most capable US open-weight model to date" is Meta's, and the card's own rows show why the geographic qualifier does the work: several non-US open models score higher on coding and agentic tests. All numbers are Meta's, from harnesses Meta chose.
The license, read plainly
The weights ship under the Muse Model License, not an OSI-approved open source license. The grant is broad: use, copy, modify, fine-tune, distill, merge, publish, distribute, sublicense and sell. Two conditions stand out. Products using the model or its derivatives must display "Built with Muse Spark Open". And if you offer models as a service and your group revenue passes 20 million US dollars in any twelve-month period, you must notify Meta and sign a separate agreement first. A usage policy restricts some uses. "Open-weight" is the accurate description, and for most individual developers and small companies the terms are effectively free.
Safety, by Meta's own evaluations
The card reports strong refusal numbers: 99.2% of jailbreaks refused on Meta's internal cyber benchmark, 94.8% aggregated across PolyGuard's cyber subset, RedCode and CyberBench 2024, with Kimi K3 at 75.5% and 70.2% in the same setup per Meta. It reports over 98% refusal on SecureBio's BioTIER, a 98.1% honesty rate on MASK, and 0% misbehavior on the Agentic Misalignment benchmark. Meta says the model got extra safety post-training because the weights are open, and that it met the thresholds of Meta's Superintelligence Scaling Framework. The full evaluation report was not out yet when I checked, so every safety number here is Meta's own and unaudited.
Analysis: my reading
The desktop threshold is the real story. A model in this class running on two machines under a desk changes the privacy and cost math for agentic coding work: the code stays in the room. The honest reading of the card is less triumphal. The model is competitive, not dominant, and its clearest leads are on safety-refusal rows that Meta both designed and measured. The license keeps a hook in exactly one group: large model-as-a-service providers. Everyone else gets terms close to Apache-style freedom with an attribution string.
What I could not confirm
I did not download or run the model. The Mac Studio demos are Meta's videos and claims, with no published throughput numbers in the sources I read. The benchmark harnesses are named per row but the full evaluation methodology report was still pending. I did not read the usage policy's full restriction list, only the license text quoted on the model card.
A check you can run
The hybrid attention claim is arithmetic you can redo. Take 16 KV heads with head dimension 64 in BF16, 62 layers, 16 of them full attention and 46 sliding over a 2048-token window. The code below computes the KV cache at 1M tokens both ways. It lands near a quarter, which is what the vLLM post says. If you plan to serve the model, run the same math with your own context length before you buy anything.
Illustrative estimator, not Meta's code
# Illustrative estimator, not Meta's code. Hybrid-attention KV cache at 1M tokens.
layers_total = 62
full_layers = 16 # vLLM day-0 post: 46 sliding-window + 16 full
window = 2048
tokens = 1_000_000
kv_per_token_layer = 16 * 64 * 2 * 2 # 16 KV heads, head dim 64, K+V, BF16
all_full = layers_total * tokens * kv_per_token_layer
hybrid = (full_layers * tokens + (layers_total - full_layers) * window) * kv_per_token_layer
print(f"all-full: {all_full/2**30:.0f} GiB hybrid: {hybrid/2**30:.0f} GiB ratio: {hybrid/all_full:.2f}")
# -> all-full: 236 GiB hybrid: 61 GiB ratio: 0.26 (vLLM says "just over a quarter")