Skip to content
FrontierDot

Meta Muse Spark Open: 816B Weights on Hugging Face, and the Two-Mac-Studio Claim

Updated

Live timeline · last updated · 2 entries in the update log

Meta says this is the most capable US open-weight model to date and that two Mac Studios can run it. The pages let us check the second claim in part, and the first not at all.

On Oct 15 Meta released the weights of Muse Spark Open, an 816B-total, 42B-active multimodal model with a 1M-token context under a custom Muse Model License; on Oct 22 Paradigm, Meta and Andromeda opened a $100,000 H200 kernel-and-serving challenge around those weights.

Read on 2026-10-17 for the weight release, and again on 2026-10-24 (UTC+8) for the kernel challenge: Meta’s announcement, the Hugging Face model card and launch post, the Transformers documentation page, Paradigm’s challenge post, openkernelchallenge.dev, and Daniel Gross’s Oct 22 X post. Benchmark charts on Meta’s page are images we could not read as text, so we report no scores. We did not run the model.

Timeline

  1. Meta releases Muse Glimmer, a small dense model in the same Muse series, according to Hugging Face’s launch post for the later model. The post says Glimmer is ideal for running locally and shares a tokenizer and vocabulary with Muse Spark Open. We found no official page giving a day for this release, so the node carries only the month.

    Hugging Face: Welcome Muse Spark Open

  2. Muse Spark Open support is contributed to Hugging Face Transformers, according to the Transformers documentation page for the model, a day before the weights were announced. The page describes a 62-layer language backbone routing each token to 8 of 256 experts, with BF16 and NVFP4 checkpoints.

    Transformers docs: Muse Spark Open

  3. Meta announces the release. Muse Spark Open has 816B total parameters, 42B active, and a 1M-token context window, built for coding and agentic work. Weights are offered on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, with hosted APIs at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The license is the Muse Model License, which Meta describes as having limited commercial terms.

    Meta AI Research: Introducing Muse Spark Open

  4. Paradigm, working with Meta and Andromeda, launches the Muse Spark Open Kernel Challenge: a $100,000 prize pool to make Muse Spark Open run faster on NVIDIA H200 GPUs. The contest has two tracks that run in series and conclude on November 19 — first five kernel problems against an SGLang baseline, then a serving track that modifies SGLang on a full 8×H200 node. First place receives $50,000 plus an H200 signed by the Meta team. Submissions are licensed under Apache 2.0.

    Paradigm: Muse Spark Open Kernel Challenge

What was released, on which day

Meta’s announcement is dated October 15, 2026, and the Hugging Face model card gives the same release date. The model card states the essentials in a few lines: 816B total parameters, 42B active, a context window of 1M tokens, natively multimodal, built for coding and agentic work, license named the Muse Model License. The authors are listed as Meta Superintelligence Labs.

One detail on the model card is easy to miss and worth quoting in spirit. The open-weights version has additional safety post-training intended to make it more robust against misuse as a standalone model without system guardrails, and Meta recommends deploying it as part of a broader system with guardrails suited to the use case. The PyTorch blog post adds that this open version shares the same base model as Muse Spark 1.3. That ties the release to a product line that existed before the weights did.

Meta’s page also says the model met the thresholds its Meta Superintelligence Scaling Framework requires for release, that detailed results are in the model card, and that a full evaluation report would be available soon. At the time we read it, that report was a promise, not a link.

The architecture, as documented

The Hugging Face launch post gives the most detail. The language model has 62 layers and a hidden dimension of 8192, with 128 query heads and 16 key-value heads. Each mixture-of-experts block has 256 experts and a router selecting 8 per token. The post says the experts hold about 97% of the total parameters, and that the model also has a 1.8B-parameter perception encoder for images and video frames and a convolutional audio encoder that emits one token every 80 milliseconds, or 12.5 tokens per second. It accepts text, images, audio and video and generates text.

The post makes the same point about memory that applies to every large mixture-of-experts model: only 8 experts run per token, so each token is computed with about 42B parameters, but all 256 experts stay in memory. Deployment needs the memory of an 816B model and runs at the speed of a 42B one. It also works out the cache: about 250 KB per token in BF16, or about 266 GB for a full one-million-token context, on top of the weights.

The checkpoint formats are BF16 and NVFP4. The post says the BF16 checkpoint is about 1.5 TB, and that in Transformers it runs across four nodes of eight H200 GPUs, while the NVFP4 version runs on a single B200 node. The model card is blunter about the floor: a GPU setup with at least 700 GB of VRAM for NVFP4 and 2 TB for BF16.

The desktop claim

Meta’s headline for the release is that it brings leading agentic capabilities to your desktop for the first time, on two Apple M5 Ultra Mac Studios using the latest MLX framework. The page explains the arithmetic it relies on: each Mac Studio’s unified memory lets the integrated GPU use up to 512 GB, so two machines can hold the whole model, and the chip’s 1.2 TB/s of memory bandwidth is, in Meta’s words, fast enough to read the small slice each token needs.

What can be checked from the Hugging Face post is a different, smaller setup. Its authors tested quantized versions on a Mac M3 Ultra with 512 GB. A 2-bit quant through llama.cpp used 304 GB and ran at 18 tokens per second. A 3-bit MLX quant took about 344 GB and ran at about 225 tokens per second on prefill and 21 on generation for a 1024-token prompt and 128 generated tokens. The same post notes that even the 3-bit quantization needs more than 256 GB of RAM.

Those are real numbers from a different machine than the one in Meta’s headline, and they show what compromise looks like: aggressive quantization, one very large Mac, and speeds that are usable but modest. We found no measurement on the pages for the two-M5-Ultra configuration itself. The claim may well hold, but the checkable evidence is for the M3 Ultra.

License and availability

Meta says it is releasing the weights under the Muse Model License “with limited commercial terms, which allow many developers to use the model for free.” That is not Apache 2.0 or MIT, and we have not summarized the license text, so read it before building a product. The distribution list is unusually broad for a launch: Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models for weights, and seven API providers for hosted access.

Inference support is also wide on paper. The model card names SGLang, vLLM, Hugging Face, TensorRT-LLM, Unsloth and llama.cpp, and the launch post shows serving recipes for llama.cpp, SGLang and vLLM. One caveat from those recipes: the vLLM reference integration supports text and image inputs with the NVFP4 checkpoint and requires Blackwell-series GPUs, so full multimodality is not the same in every engine.

What we could not verify

  • “Most capable US open-weight model to date.” It is Meta’s own label, repeated by Hugging Face’s post. Meta’s benchmark charts are images, so we did not read or reproduce any scores.
  • The two-M5-Ultra Mac Studio setup. We found measurements only for an M3 Ultra.
  • The safety evaluation. Meta says a full report is coming; we did not see it.

Our take

The release is substantial and well supported at launch. The superlatives are Meta’s, and the license is a custom one, so the practical questions for a developer are the usual two: can I afford the memory, and can I ship with this license. For most people the answer to the first is a hosted API, which Meta lists, rather than a desk. The October 22 kernel challenge is a separate bet: that open weights become more useful when the community can push Hopper serving past the SGLang baseline. This page is a live timeline. When the evaluation report, challenge results, new quantizations or license clarifications appear, they get a dated node and an update log entry, and the one-line conclusion at the top changes with them.

For the other large open-weight release of the same week, with a permissive license and a different hardware story, see the Reflection Beam timeline.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

When did Meta release Muse Spark Open?

October 15, 2026, per Meta’s announcement and the Hugging Face model card.

What license does Muse Spark Open use?

The Muse Model License, which Meta describes as having limited commercial terms. It is a custom license, so read the text before commercial use.

Can Muse Spark Open run on a Mac?

Meta says it runs on two Apple M5 Ultra Mac Studios with MLX. Hugging Face’s tests used a 512 GB M3 Ultra with 2-bit and 3-bit quantized versions.

Update log

  • — Added the Oct 22, 2026 Muse Spark Open Kernel Challenge node from Paradigm, Meta and Andromeda.
  • — First published.

Sources