Image: Meta (official) — header from Meta's "Build with Muse Spark Open" developer blog (Oct 15, 2026)
On October 15, 2026, Meta released the weights for Muse Spark Open, the open-weight version of its Muse Spark model. Meta calls it "the most capable US open-weight model to date." The headline numbers are big: 816 billion total parameters, about 42 billion active per token, a 1-million-token context window, and native input for text, images, video, and audio. It is built for coding and agent work, meaning tasks where the model reads a codebase, plans a change, calls tools, checks the result, and fixes its own mistakes.
"Open weights" is the part that gets attention, but most people will not download a model this size on day one. So this guide is organized by what you actually have: a credit card and an API key, a serious GPU server, or a stack of Mac Studios. Everything below comes from Meta's announcement, Meta's developer blog, the Hugging Face model card, and OpenRouter's public model listing.
1. Decide which path fits you
- Just want to test it: use a hosted API. Meta lists Baseten, Fireworks AI, Groq (limited preview), Modal, Nebius, OpenRouter, and Together AI as launch partners. No hardware needed.
- Have a GPU cluster or cloud budget: serve it yourself with SGLang, vLLM, or TensorRT-LLM, or pull it from Amazon SageMaker AI, Google Cloud Model Garden, or Microsoft Foundry Models.
- Want it on your desk: the MLX route on Apple M5 Ultra Mac Studios, or Unsloth's compressed GGUF files on a large workstation.
My advice: start with the hosted route even if you plan to self-host. A few dollars of API calls will tell you whether the model handles your tasks before you commit hundreds of gigabytes of memory to it.
2. The quickest test: a hosted API
OpenRouter is the easiest single door because one key reaches many providers. Its listing uses the model ID meta/muse-spark-open, shows a 1,048,576-token context, and accepts text, image, file, audio, and video input with text output. At the time of writing, OpenRouter listed $1.25 per million input tokens and $4.25 per million output tokens. Prices change, so check the listing before you run anything big.
The steps:
- Create an OpenRouter account, add credit, and generate an API key.
- Send a request to OpenRouter's OpenAI-compatible chat completions endpoint with
"model": "meta/muse-spark-open". - Set the reasoning effort. OpenRouter marks reasoning as mandatory for this model, so you cannot switch thinking off, only turn it down.
If you already use Fireworks AI or Together AI, they are also launch partners, and Meta notes both offer managed fine-tuning on top of inference.
3. Set the two settings that matter
The model card is unusually direct about tuning, and getting it right saves both money and frustration.
- Reasoning effort: accepts
low,medium,high, orxhigh. The default isxhigh, the most expensive setting. Meta recommends high or xhigh for hard problems, coding, and agent tasks, and low or medium for faster, shorter answers. Any other value returns an error. On a self-hosted server you can pass it asreasoning_effortor throughchat_template_kwargs. - Sampling: temperature 1.0, top_p 1.0, and top_k disabled. Do not copy the lower temperatures you may use with other models; Meta ran its own benchmarks with these settings.
A practical pattern: run quick lookups and summaries at medium, and only move to xhigh for multi-step coding or agent runs where a wrong step costs more than extra tokens.
4. Self-hosting on GPUs
Before you start, request access. The Hugging Face repositories (meta-models/Muse-Spark-Open-816B-A42B and the -NVFP4 version) are gated with manual approval, so you need to accept the license and wait.
Then size your hardware honestly. Meta's developer blog lists the full BF16 weights at about 1.6 TB and the NVFP4 version at about 530 GB, serving on four Blackwell Ultra GPUs or eight H200s. The model card is more conservative, saying you need at least 700 GB of GPU memory for NVFP4 and 2 TB for BF16. Plan for the bigger number, because the long context window also eats memory.
Meta's blog gives a vLLM starting point that looks like this:
vllm serve meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \ --tensor-parallel-size 4
Its full example also adds a speculative decoding config that uses a small draft model called DSpark (Inferact/Muse-Spark-Open-DSpark). A draft model guesses several tokens ahead and the big model checks them in one pass, which speeds up output. SGLang has its own draft model, DFlash V2 (RadixArk/Muse-Spark-Open-DFlash2). Both servers expose an OpenAI-compatible API, so existing client code usually only needs a new base URL and model name.
5. Running it on your desk
This is the part Meta and Apple promoted together. MLX, Apple's machine-learning framework, can run the NVFP4 checkpoint split across two Mac Studios with M5 Ultra and 512 GB of unified memory each, or four with 256 GB each, for about 1 TB in total. The machines connect directly over Thunderbolt 5 with RDMA, and every layer is split across them, so each Mac holds half the weights and does half the math for each token. Meta points to a "Muse Spark Open on Mac" cookbook and an MLX adapter for the exact steps.
The other desk option is Unsloth. Its Dynamic GGUF files compress the model down to about 2-bit, and Meta's blog says the smallest version fits in roughly 300 GB of combined RAM and VRAM, or unified memory on a Mac. Unsloth Desktop runs on macOS, Windows, and Linux, can push layers into system RAM when the GPU runs out, and can serve OpenAI- and Anthropic-compatible endpoints for your coding agents. Expect some quality loss at 2-bit, and test against the hosted version before trusting it.
6. Read the license before you build a business on it
Muse Spark Open ships under the Muse Model License, not a standard open-source license. Meta describes it as having "limited commercial terms" that let many developers use the model for free. Two clauses are worth knowing. First, any product or service that uses the model, or a model fine-tuned or distilled from it, must prominently show "Built with Muse Spark Open" in its interface, documentation, website, or settings. Second, if your company offers AI models to others through an API (Model as a Service) and its total gross revenue passes $20 million in any 12-month period, you must notify Meta and sign a separate agreement before offering this model that way. Internal use, research, and using the model to power features of your own product that is not a model-hosting service are outside that second clause. Meta also claims no ownership of outputs.
Meta also recommends deploying the model inside a larger system with your own guardrails, especially once it is calling tools on someone's behalf.
My take
The interesting part is not that a huge model is open. It is that Meta shipped it with day-one support across almost every serving stack, with draft models and quantized files ready. For most readers, the useful move this week is a cheap hosted test at medium effort on a real task from your own work, such as a code review or a long document. If it beats what you use now, then look at self-hosting. The two-Mac-Studio setup is impressive, but it is a large purchase to make before you know the model fits your work.
Sources
- Meta AI Research: Introducing Muse Spark Open (Oct 15, 2026)
- Meta for Developers: Build with Muse Spark Open
- Hugging Face: Muse Spark Open model card
- Muse Model License
- OpenRouter: Meta Muse Spark Open listing
댓글
댓글 쓰기