Meta’s Muse Spark Open: what is in the release
From Meta AI Research blog: Introducing Muse Spark Open · Open at Meta AI Research blog: Introducing Muse Spark Open
On 15 October 2026 Meta released the weights of Muse Spark Open, a mixture-of-experts model that reads text, images, video and audio and writes text. Meta calls it "the most capable US open-weight model to date". We cannot check that claim, and this page does not repeat it as fact. Below is what Meta's post, the model card and the Hugging Face blog confirm, what the licence asks of you, and what the sources leave out.
What Meta confirms about Muse Spark Open
These points come from Meta's announcement, the model card on Hugging Face and the Hugging Face blog post, all dated 15 October 2026.
| Item | What the pages say | Where |
|---|---|---|
| Size | 816B parameters in total, 42B active per token | Announcement, model card |
| Design | Mixture of experts: 256 experts per layer, 8 used for each token, 62 layers | Model card, Hugging Face blog |
| Inputs and output | Input: text, image, video and audio. Output: text | Model card |
| Context | 1M tokens at inference | Model card |
| Languages | Trained on data from more than 100 languages | Model card |
| Weight formats | BF16 and NVFP4 (4-bit). The BF16 checkpoint is about 1.5 TB | Model card, Hugging Face blog |
| Where to get the weights | Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models | Announcement |
| Hosted API | Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI | Announcement |
| Licence | Muse Model License, which Meta says has "limited commercial terms" | Announcement, model card |
Meta says the model is "built for coding and agentic work". The model card lists uses such as multi-step reasoning, coding agents, tool use and reading screenshots, charts, documents, audio and video.
When the weights became public
All three sources give 15 October 2026 as the release date. The model card's overview table says "October 15th, 2026", and the licence text on the same page says "Release date: October 15, 2026". Meta's post is dated October 15, 2026 and says the model "is available today". The Hugging Face post was published the same day.
The repository itself is older. The Hugging Face record for the repository shows it was created on 15 September 2026 and last changed on 15 October 2026. None of the sources say whether the weights could be downloaded before 15 October. Access is also gated: the model card asks you to log in, give your name, date of birth, country and organization, and accept the licence before you can download.
What the licence asks of you
"Open weights" here does not mean "no terms". The Muse Model License on the model card allows you to use, change, fine-tune, distill and sell copies of the model, under these conditions:
| Condition | What the licence says | Where |
|---|---|---|
| Credit | Show "Built with Muse Spark Open" in the product, or in its documentation, website or settings menu. This covers models trained or distilled from it, or from its outputs | Model card |
| Large API sellers | If you offer AI models to others through an API and the gross revenue of you and your affiliates is over 20 million US dollars in any 12-month period, you must notify Meta and sign a separate agreement before you offer it | Model card |
| Who that rule skips | Internal use, research, and using the model to power features of your own product that is not a model service | Model card |
| Usage Policy | Use must follow the attached Usage Policy, which lists banned uses | Model card |
| Age | "Muse Spark Open is not intended for individuals under the age of 18" | Model card |
| Outputs | "Meta claims no ownership of outputs generated by the Software" | Model card |
The Hugging Face blog post does not discuss the licence. Our piece on open-weight and paid models explains why the licence matters as much as the weights.
What it takes to run it
The model is large. The sources give these setups. The speeds are what the Hugging Face team says it measured on Apple machines. No source gives serving speeds for the GPU setups.
| Setup | What the source says | Where |
|---|---|---|
| Memory need | At least 700GB of GPU memory for NVFP4 and 2TB for BF16 | Model card |
| Transformers, BF16 | 4 nodes of 8 H200 GPUs | Hugging Face blog |
| Transformers, NVFP4 | One B200 node | Hugging Face blog |
| vLLM reference setup | NVFP4 only, needs Blackwell GPUs, text and image input, 128K-token context | Hugging Face blog |
| SGLang reference setup | NVFP4 on four GB300 GPUs | Hugging Face blog |
| Hugging Face Inference Endpoints | NVFP4 on 8 RTX Pro 6000 (98 GB) with 131,072 tokens of context, or 8 H200 (141 GB) for 1,048,576 tokens | Hugging Face blog |
| llama.cpp, 2-bit | M3 Ultra Mac with 512GB: 304 GB used by the model, 18 tokens per second | Hugging Face blog |
| MLX, 3-bit | M3 Ultra with 512 GB: about 344 GB, about 21 tokens per second when writing and 225 when reading the prompt | Hugging Face blog |
| Two Mac Studios | Meta says the model runs on two Mac Studios with M5 Ultra, each with up to 512 GB of memory for the GPU | Announcement |
Two points to keep in mind. The 1M-token context is the maximum. The vLLM reference setup uses 128K tokens, and the Hugging Face blog says a full 1M-token context needs about 266 GB of memory for the cache alone in BF16. And the small versions that fit on one Mac are not the checkpoints Meta tested in its benchmark table. See what it takes to run an open AI model yourself for the cost beyond the hardware.
What Meta's benchmark table shows
Meta's post has a "Key Benchmarks" heading but no numbers in its text. It points to an evaluation report and to the model card. The model card has a table that compares the model with other open-weight models. Meta says it ran its own model at "xhigh reasoning effort". Below are some rows, with Kimi K3 as one comparison, because Kimi K3 has a figure on most rows.
| Test (as named by Meta) | Muse Spark Open, Meta says | Kimi K3 (max), in Meta's table | Where |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 39 | 44 | Model card |
| MCP-Atlas | 89.2 | 82.3 | Model card |
| OSWorld Verified | 80.4 | 84.8 | Model card |
| Toolathlon Verified | 72.8 | 76.5 | Model card |
| HLE (with tools) | 61.0 | 56.0 | Model card |
| GPQA Diamond | 89.0 | 93.5 | Model card |
| DeepSWE 1.1 | 60 | 68 | Model card |
| Terminal-Bench 2.1 | 76.4 | 85.0 | Model card |
| VideoMMEv2 | 61.5 | 65.4 | Model card |
Meta's own table does not show one model ahead everywhere. On these rows Muse Spark Open is ahead of Kimi K3 on two and behind on seven. The table also names Inkling, Nemotron 3 Ultra, Mistral Large 4, Qwen3.8 and GLM 5.3, with many empty cells.
The model card also reports safety tests. Meta says the model refused 99.2% of jailbreaks on its internal cyber misuse benchmark and 94.8% across three public ones, with no system safeguards. Meta's post says it "met the thresholds our framework requires for release".
What is still missing
- The full evaluation report. Meta's post says "a full evaluation report" is "available soon". The model card links "our report", which we have not reviewed for this page.
- Test details for the comparison. The model card gives the settings it used for Muse Spark Open. For the other models it shows only a label such as "max" or "xhigh". It does not say how many examples were used, whether the other scores were run by Meta or copied from elsewhere, or what the column "Beam" refers to.
- When the weights became downloadable. The repository record dates from 15 September 2026. The sources give 15 October as the release date and do not say more.
- GPU serving speeds. None of the three sources (Meta, model card, Hugging Face blog) gives tokens per second for serving the model on GPUs.
What this means if you run AI work
A model with public weights gives you a choice of where it runs and what it costs per request. This one needs a lot of memory, and its licence has conditions. That is a reason to test it, not yet a reason to switch.
- Run your own evaluation. Use your own tasks and pass mark, on the version you would actually run.
- Read vendor numbers as claims. Meta's table is Meta's. Wait for the full report before you rely on it.
- Check the licence. The credit line, the rule for large API sellers and the Usage Policy all apply.
We track how model prices and capabilities change in Model Watch.
Starfield is not affiliated with Meta. This article is general information and not investment advice.
Related
Model Watch · Open-weight or paid AI models: what is the difference? · What it takes to run an open AI model yourself