A few days after Meta’s closed-weight Muse Wonder, the company released an open-weight model. Muse Spark Open came out October 15. This post draws on Meta’s announcement, TechCrunch and an OrcaRouter analysis of the license and benchmarks. I haven’t downloaded or run it. Meta’s claims are labeled as Meta’s.

The specs

  • 816B total parameters, 42B active per token, a sparse mixture of experts. OrcaRouter lists 62 layers, 256 experts per MoE layer and 8 selected per token.
  • 1M-token context window (1,048,576), shared across text, image, video and audio. OrcaRouter says audio and video are input-only and output is text.
  • Meta describes it as built for coding and agentic work, and “the most capable US open-weight model to date,” which is Meta’s own claim.
  • Meta also says it applied extra safety post-training because it’s open-weight and that it met release thresholds under its scaling framework. The full evaluation report is “available soon,” so that claim can’t be checked yet.

Where you can get it

According to Meta: weights on Hugging Face, plus Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry. Hosted APIs include Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together. Inference engines listed include SGLang, vLLM, TensorRT-LLM, Transformers, Apple MLX, llama.cpp and Unsloth.

The hardware bill

This isn’t a laptop model.

  • Meta highlights a setup of two Apple M5 Ultra Mac Studios, each with up to 512 GB of unified memory.
  • TechCrunch, citing Engadget, says local use requires at least 1 TB of RAM, which is consistent with the two-machine demo.
  • OrcaRouter puts BF16 weights at roughly 1.5 TB and the KV cache at full 1M context at about 266 GB, and says an NVFP4 version is said to run on a single B200 node. Meta’s page specifies no quantization formats, so treat that as unconfirmed.

“Open” has fine print

OrcaRouter summarizes the Muse Model License this way, so verify against Meta’s text:

  • Service threshold: offering third parties access to the model as a product or API with more than $20 million in gross revenue over twelve months requires a separate agreement. Building it into your own product is exempt.
  • Attribution: products using it must display “Built with Muse Spark Open.”
  • Allowed: fine-tuning, distillation, merging, publishing, sublicensing and selling derivatives, with no Meta ownership of outputs.
  • Gated repo: the request form asks for legal name, date of birth and organization details.

So it isn’t MIT or Apache. For a hobbyist or startup, probably fine. For a company reselling inference, read the agreement. I covered a permissively licensed small model in the Mellum 2.1 post if license simplicity matters more than size.

Benchmarks: one independent, the rest Meta’s

  • Independent (Artificial Analysis): Intelligence Index 39, ranked 9th of 118 open-weight models above 150B parameters, where the class median is 19. About 103 output tokens per second, and roughly $0.84 per completed index task on one hosted route.
  • Meta’s scorecard: eight benchmarks run in Meta’s own agent harness, including Terminal-Bench 2.1 at 76.4 and MCP-Atlas at 89.2. Not reproduced by anyone I found.
  • Gaps: no public audio or video evaluation, even though those are headline inputs, and reasoning can’t be turned off.

One search result, a PyTorch blog post, is titled “Muse Spark 1.2 Is Now Open Weights,” so the naming differs between sources. I haven’t opened it.

My take

The interesting part is the combination: an agent-focused model with a 1M window, in your hands, on hosted APIs from seven providers. If you want to test it without buying two Mac Studios, use one of those APIs and compare against your current model on your own repo, as I suggested in the model comparison post.

Sources: Meta, TechCrunch, OrcaRouter.