This guide explains what these terms mean in plain language, how to read the benchmark results Meta published, and how to choose between a hosted API and running the model yourself. Details come from Meta's announcement and the model card, with independent coverage linked where it adds context.
1. Muse Spark Open at a Glance
Muse Spark Open is a natively multimodal model that Meta released on October 15, 2026, with its weights available through the Hugging Face collection. Meta says it is built for coding and agentic work, including long-horizon tasks, complex software engineering, and multi-step reasoning.
| Detail | What is reported |
|---|---|
| Release date | October 15, 2026 |
| Total parameters | 816B |
| Active parameters | 42B per token |
| Context window | 1M tokens |
| Architecture | Mixture of experts, 256 experts per layer, 8 active per token |
| Inputs and output | Text, image, video, and audio in, text out |
| Checkpoints | BF16 and NVFP4 |
| License | Muse Model License, with limited commercial terms |
| Access | Hosted APIs and downloadable weights |
Meta also calls it the most capable US open-weight model to date. That is the company's own description, so it is worth comparing against your own tests.
2. What Is an Open-Weight Model?
An open-weight model is an AI model whose trained weights are published, so that others can download it, run it on their own infrastructure, and often fine-tune it for their own needs.
Weights are the numbers a model learns during training. They hold what the model has learned, and they are what you need to run the model without relying on the original developer's servers.
2.1 How is open-weight different from open source?
The two terms are often mixed up, but they are not identical. Open weights mean the trained model is available. They do not always include training data, full training code, or unrestricted usage rights.
| Question | Open weights in general | Muse Spark Open |
|---|---|---|
| Are trained weights available? | Yes | Yes, with access requested through Hugging Face |
| Is the license standard? | Varies by release | Custom Muse Model License |
| Is training data included? | Usually not | Not part of the announcement |
| Can it be fine-tuned? | Often | Yes, and Meta points to PyTorch's TorchTitan for tuning |
2.2 Why do open weights matter?
- Deployment control: you choose where the model runs, from a hosted provider to your own hardware
- Data handling: running the model in your own environment can simplify data governance
- Customization: the model can be tuned for a specific domain or workflow
- Provider choice: the same model can be offered by several platforms, which gives teams options
3. What Does 816B Total and 42B Active Mean?
Parameters are the learned values inside a model. Muse Spark Open has 816B total parameters (816 billion), with 42B active for each token.
The short answer: total parameters describe the size of the whole model, while active parameters describe how much of it does the work on each token. Here, that is roughly 5 percent of the model at any one step.
3.1 How does mixture of experts work?
According to the model card, Muse Spark Open is a mixture of experts model. It contains many specialized sub-networks called experts, and a routing step picks only a few of them for each token. The rest stay idle for that step.
| Number | What it describes | What it affects |
|---|---|---|
| 816B total parameters | The full model | Memory needed to hold it |
| 42B active parameters | The share used per token | Compute and speed per token |
| 256 experts per layer | The pool of specialists | Capacity to cover varied tasks |
| 8 active experts per token | The share chosen by the router | How much work each token triggers |
3.2 How much memory does it need?
Memory depends on total parameters, because the whole model has to be stored where it runs. Speed depends mostly on active parameters, so each token costs far less than the full model size suggests. That is why the memory figures below are large even though the model can respond quickly.
| Setup | Reported memory figure | Where it comes from |
|---|---|---|
| BF16 checkpoint | At least 2 TB | Model card |
| NVFP4 checkpoint | At least 700 GB of VRAM | Model card |
| Two M5 Ultra Mac Studios with MLX | 512 GB each, about 1 TB combined | Meta's announcement |
| Smallest community 2-bit quantization | About 300 GB of combined RAM and VRAM | Reported by Developers Digest |
Lower precision versions reduce memory but trade away some quality and speed. For a standard laptop or a single gaming PC, hosted access is the practical route.
4. What Does a 1M Token Context Window Allow?
A context window is the amount of text a model can consider at one time. It includes your instructions, any files you provide, the conversation so far, and the model's own output.
Tokens are small chunks of text. In English, one token is roughly three quarters of a word, so 1M tokens is very roughly 750,000 words. Code and other languages tokenize differently, so treat this as an estimate.
| Use case | What it looks like in practice |
|---|---|
| Large codebases | Reading many files together for refactoring or finding related code |
| Long documents | Reviewing contracts, reports, or research papers in one request |
| Agent sessions | Keeping tool outputs, logs, and intermediate steps in view over long tasks |
| Multi-source comparison | Reviewing several documents side by side without splitting them up |
A larger window does not remove every limit. Longer inputs use more memory and take more time to process, and quality on very long inputs can vary by task. Many teams still pair long context with retrieval, so the model sees the most relevant material rather than everything.
5. What Did Meta Test on Desktop Hardware?
Meta says it tested agentic workflows running entirely on-device, using Muse Code powered by Muse Spark Open. The setup was Apple Mac Studio with M5 Ultra and the latest version of MLX, with two machines holding the full model.
| Demonstration | What the model did | What it shows |
|---|---|---|
| Software development | Shipped real iOS changes to Ice Cubes, an open-source Mastodon client, including bug fixes, new features, and a full build | Multi-step coding work from task to finished build |
| Video editing | Drove DaVinci Resolve end to end, reviewing footage frame by frame to find the strongest moments and crop them vertically | Visual understanding combined with tool control |
These are demonstrations shared by Meta, not independent tests, and they were run on a high-end setup. They are best read as examples of the kind of long, multi-step work the model is designed for.
6. How to Read the Muse Spark Open Benchmark Table
Benchmarks are standardized tests used to compare models on the same tasks. Meta compares Muse Spark Open with Inkling, Nemotron 3 Ultra, Beam, Mistral Large 4, Qwen3.8-2.4T-A95B, GLM 5.3, and Kimi K3, across agentic, knowledge and reasoning, coding, and multimodal tests.
6.1 Where does it lead and where does it trail?
| Category | Benchmark | Muse Spark Open | Highest score in table | Leader |
|---|---|---|---|---|
| Agentic | JobBench | 61.8 | 61.8 | Muse Spark Open |
| Agentic | MCP-Atlas | 89.2 | 89.2 | Muse Spark Open |
| Agentic | GDPval-AA v2 | 1401 | 1653 | GLM 5.3 |
| Knowledge and reasoning | HLE-Diamond (no tools) | 26.1 | 26.1 | Muse Spark Open |
| Knowledge and reasoning | SciCode | 58.5 | 59.5 | Kimi K3 |
| Knowledge and reasoning | AA-Omniscience | 25 | 25 | Muse Spark Open |
| Coding | Terminal-Bench 2.1 | 76.4 | 85.0 | Kimi K3 |
| Coding | DeepSWE 1.1 | 60 | 68 | Kimi K3 |
| Multimodal with tools | ZeroBench | 50.0 | 50.0 | Muse Spark Open |
| Multimodal with tools | BabyVision | 80.2 | 85.7 | Kimi K3 |
| Multimodal with tools | SimpleVQA | 73.6 | 73.6 | Muse Spark Open |
The pattern is mixed. Muse Spark Open has the top score in six of the eleven rows, mostly in agentic, knowledge, and multimodal tests. In the coding rows, Kimi K3 and GLM 5.3 score higher. Teams with coding focused needs may want to test it on their own tasks before deciding.
6.2 Why do test setups matter?
A score is only comparable when the conditions are comparable. A few details in the table are worth noticing:
| Detail | Where it shows up | Why it matters |
|---|---|---|
| Reasoning settings | Columns marked xhigh, max, or thinking | Models were run at different effort levels |
| Agent harness | Terminus 2 agent on Terminal-Bench 2.1, and a best available or self-reported harness on DeepSWE 1.1 | The tooling around a model can change results |
| Different scales | GDPval-AA v2 values in the thousands, and a negative value on AA-Omniscience | Compare scores only within the same row |
| Missing results | Dashes and "Not supported" entries | Not every model was tested on every benchmark |
The table is also published by Meta, so independent evaluations and your own tests on real workloads give a fuller picture. Meta's research announcement and the model card have more detail.
7. API vs Local Deployment: Which Route Fits Your Team?
There are two main ways to use Muse Spark Open. The right one depends on your data needs, usage volume, and available hardware.
| Factor | Hosted API | Self-hosted or local |
|---|---|---|
| Setup effort | Low | High |
| Hardware | None to buy | Significant memory required |
| Cost model | Pay by usage | Hardware and operations cost |
| Data control | Data goes to a provider | Data stays in your environment |
| Customization | Limited to what the provider offers | Full tuning and configuration |
| Best for | Evaluation, variable usage | Strict data needs, steady high volume |
7.1 Hosted API options
Meta lists API access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Availability and terms vary by provider, so check each one for pricing and data policies.
7.2 Self-hosted options
Weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Meta lists support for SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon.
The release also includes custom, auto-tunable Helion GPU kernels in SGLang. Meta reports kernel-level speedups of 1.03 to 3.95 times on NVIDIA and AMD GPUs, and a 3 percent end-to-end speedup on AMD MI350. Serving commands for each engine are in Meta's developer post.
7.3 Which route fits which team?
| Team type | Practical starting point |
|---|---|
| Startup shipping an agent product | Hosted API, then reassess as volume grows |
| Enterprise with strict data rules | Self-hosting or a private cloud deployment |
| Research or fine-tuning group | Downloaded weights with TorchTitan for tuning |
| Company offering the model to others as a service | Read the license terms before building |
For most teams, a hosted API trial is the simplest way to test quality and cost on real tasks. Moving to self-hosting makes more sense once data requirements, volume, or customization needs justify the hardware and operations work.
8. License, Safety, and Access Notes
The Muse Model License allows many developers to use the model for free, with some conditions. Independent coverage of the license, including The New Stack, points to a few terms worth reading before you build.
- Attribution: products using the model or a derivative should display "Built with Muse Spark Open"
- Large service providers: businesses offering the model's outputs to third parties as a service may need a separate agreement with Meta above a revenue threshold, so review the license text
- Gated access: downloading the weights from Hugging Face requires requesting access
On safety, Meta says Muse Spark Open received additional safety post-training because open weights carry their own considerations. It states the model was evaluated under the Meta Superintelligence Scaling Framework and met the thresholds required for release. Detailed results are in the model card, and a full evaluation report is expected soon.
9. Conclusion
Muse Spark Open adds another large open-weight option for coding and agentic work. Its 816B total and 42B active parameters explain both its memory needs and its speed profile, and its 1M token context window suits long, multi-step tasks.
In Meta's published benchmarks it leads several rows and trails in others, particularly in coding. That makes testing on your own workloads the most reliable next step. A hosted API trial is a low-effort way to start, and the downloadable weights and range of inference engines leave room to move to your own infrastructure later.