Meta released Muse Spark Open on October 15, 2026. It has 816 billion total parameters, 42 billion active per token, and a 1M-token context window. At 4-bit NVFP4 quantization, the parameters alone take about 493 GB, so the published 4-bit results use at least two M5 Ultra Mac Studios. Apple measured 15.7 to 36.7 tokens per second on those clusters, using an 8,192-token context.
Model Facts
| Item | Detail |
|---|---|
| Released | October 15, 2026 |
| Size | 816B total, 42B active (mixture of experts) |
| Context window | 1M tokens |
| Input | Natively multimodal, built for coding and agent work |
| License | Muse Model License, with limited commercial terms. It is not Apache or MIT |
| Hosted APIs | Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI |
| Weights | Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Models |
| Local engines named by Meta | Apple MLX, llama.cpp, Unsloth, plus SGLang and vLLM for servers |
Memory Needed
Apple's MLX team reports that 4-bit NVFP4 shrinks the model to about 493 GB, and that native BF16 precision would need more than 1.6 TB. A Mac Studio tops out at 512 GB of unified memory with 1.2 TB/s of bandwidth. That looks like enough for 493 GB, but the parameters are not the whole workload, because activations and context cache need memory too. That is why the 4-bit tests are split across machines. Engadget, as cited by TechCrunch, gives at least 1 TB of RAM as the practical floor.
Measured Mac Studio Speed
These are Apple's published decode results, all at NVFP4 4-bit with an 8,192-token context.
| Setup | Split | Decode speed | Time to first token |
|---|---|---|---|
| 2 M5 Ultra Mac Studios | Pipeline | 15.74 tokens/s | 8.4 s |
| 2 M5 Ultra Mac Studios | Tensor | 24.99 tokens/s | 8.2 s |
| 4 M5 Ultra Mac Studios | Tensor | 36.73 tokens/s | 4.8 s |
Apple's table does not state the memory fitted to each machine. Secondary coverage describes 512 GB units for the two-Mac cluster and 256 GB units for the four-Mac cluster, and puts the two-Mac cost near $38,000, a figure it attributes to Engadget. Lower-bit single-machine builds have also been reported. We have not confirmed one against a primary measurement, so this table does not include them.
The Context Caveat
The 8,192-token test context is less than 1 percent of the model's 1M-token window. Apple separately reports 40.6 seconds to first token at 65,536 tokens on four M5 Ultras, more than eight times the 4.8 seconds above. Long-context coding sessions, the use Meta highlights, will not feel like the headline number. For how to judge these speeds, see our tokens per second benchmarks.
How Capable Is It?
Meta calls it "the most capable US open-weight model to date." That is a claim about US models, not all open models. Artificial Analysis reportedly gives it an Intelligence Index of 39, ranking 11th of 121, behind GLM 5.3 at 45 and Kimi K3 at 44 in one benchmark listing. Meta's Terminal-Bench 2.1 result is reported as 76.4. Reflection reports 80.1 for its smaller Beam on the same test. Both of those are vendor numbers. Meta said its full evaluation report would be available soon.
Should You Run It Locally?
- For privacy or research: a two-Mac tensor setup is the smallest measured configuration in Apple's report.
- For daily coding: hosted endpoints are the practical route, since local speed drops as context grows.
- For commercial products: read the Muse Model License first. Meta describes its terms as limited.
Methodology
Facts come from Meta's announcement and Apple's MLX distributed inference report, with TechCrunch for launch coverage. Presenc AI did not run these tests. Status as of October 31, 2026.
Following This With Presenc AI
Presenc AI is building tools to show how AI systems find, describe, and act on brands. Join the waitlist for access.