How to Run Muse Spark Open on Ollama: Cloud Setup Guide (2026)
Muse Spark Open is Meta's 816B-parameter coding model, cloud-only on Ollama. Run muse-spark-open:cloud, set up API access, and find true local alternatives.
Muse Spark Open is Meta Superintelligence Labs' open-weight coding and agentic model, released October 15, 2026. It's a mixture-of-experts design with 816 billion total parameters and 42 billion active per token, spread across 256 experts per layer with 8 selected on every forward pass, and it carries a 1,048,576 token context window. The weights are genuinely public on Hugging Face under the Muse Model License, but running all 816B parameters takes serious hardware. Meta's own reference setup for local inference is two Mac Studios with M5 Ultra chips. On Ollama specifically, the only listed tag is `muse-spark-open:cloud`, which routes your prompt to Meta's hosted infrastructure through Ollama's servers instead of loading anything on your machine.
That's a different situation from a model like Mistral Large 4, where no weights exist anywhere yet. With Muse Spark Open, the weights are real and downloadable, runnable through vLLM, SGLang, TensorRT-LLM, or llama.cpp if you have the hardware for it, but Ollama has not added a local pull tag, so `:cloud` is the only way to use this model through Ollama's command syntax today.
This guide covers the cloud setup end to end: installing Ollama, signing in, running `muse-spark-open:cloud` from the terminal, and generating an API key for scripts and agents. The alternatives section covers Muse Glimmer, Meta's 30B distillation of this same model that you can actually run on a single GPU or Mac, along with Kimi K2.6 and DeepSeek R1 for other cloud and local options.
Prerequisites
- Ollama 0.12 or later, installed on Linux, macOS, or Windows (no GPU or high-RAM machine required for the cloud route)
- A free account at ollama.com for the `ollama signin` step
- A stable internet connection, since inference runs on Meta's and Ollama's servers, not your hardware
- Basic terminal familiarity for running `ollama run` and `curl` commands
- (Optional) An API key from ollama.com/settings/keys if you plan to call Muse Spark Open from your own scripts or agents
Need more GPU power?
Rent a 8x H100 80GB on Vast.ai from $18.00/hr. On-demand GPU rentals by the hour, useful for running larger models without buying hardware.
In This Guide
What Muse Spark Open Is and Why Ollama Runs It in the Cloud
Muse Spark Open is Meta Superintelligence Labs' flagship open-weight model, published on Hugging Face as `meta-models/Muse-Spark-Open-816B-A42B`. It's a mixture-of-experts model with 816 billion total parameters and 42 billion active per token, built on a 62-layer architecture with an 8,192 hidden size, 128 attention heads, and 16 KV heads. Every layer routes each token to 8 of 256 available experts. Meta describes it as the most powerful open-weight model trained in the US to date, aimed at coding, tool use, and long-horizon agentic work where a model has to plan several steps, call tools, and recover when something along the way goes wrong.
| Spec | Value |
|---|---|
| Total parameters | 816 billion (MoE) |
| Active parameters | 42 billion per token |
| Experts | 256 per layer, 8 selected per token |
| Context window | 1,048,576 tokens |
| Layers | 62 |
| Hidden size | 8,192 |
| Attention heads | 128 (16 KV heads) |
| License | Muse Model License (limited commercial terms) |
| Ollama tag | `muse-spark-open:cloud` (only option) |
The model accepts text, image, audio, and video input and produces text output, and it's distributed through API providers including Together AI, Fireworks AI, Baseten, Groq, Modal, and Nebius, alongside the direct Hugging Face weights. Meta has not published formal benchmark scores for Muse Spark Open as of this guide's publish date, so treat any comparison to GLM 5.3, Kimi K3, or similar flagship models as unverified until independent evaluations appear.
Through Ollama Cloud, pricing runs $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens. That pricing applies within Ollama's own usage limits, separate from direct billing through Meta's API partners.
Set Up Ollama Cloud and Run Muse Spark Open
Running Muse Spark Open through Ollama takes three steps: install Ollama, sign in, and run the model. Nothing downloads an 816B-parameter file to your disk, since inference happens on Meta's infrastructure rather than locally.
Step 1: Install Ollama
# Linux and macOS, one-command installer
curl -fsSL https://ollama.com/install.sh | shOn Windows, download the installer from ollama.com/download, or use winget:
winget install Ollama.OllamaVerify the installation:
ollama --versionollama version is 0.12.3Step 2: Sign In to Ollama Cloud
ollama signinThis prints a sign-in URL and opens your browser. Create a free account at ollama.com, or log in if you already have one, then approve the device.
Signing in to ollama.com...
Signed in as your-usernameStep 3: Run Muse Spark Open from the Terminal
ollama run muse-spark-open:cloudOllama fetches a small manifest rather than model weights, then drops you into a prompt:
pulling manifest
pulling 4a91c0d2e7f3... 100% ââââââââââââââââââ 4.3 KB
success
>>> Send a message (/? for help)Test it with a coding prompt, since that's the model's primary design target:
>>> Write a Python function that rate-limits API calls to 10 per second using a token bucket, and explain the algorithm.The first response after signing in can take 10-20 seconds while Ollama establishes the cloud session. Responses after that stream back at normal speed.
Step 4: Verify the Model
ollama list`muse-spark-open:cloud` shows up at a few KB, not hundreds of gigabytes. That confirms it's the cloud passthrough entry, not a local download.
More how-to guides like this
Step-by-step AI guides and tool walkthroughs, straight to your inbox, twice a week.
Use Muse Spark Open in Your Own Scripts and Agents
Beyond the interactive `ollama run` session, Muse Spark Open's cloud tag works through Ollama's REST API and its built-in agent launch shortcuts, so existing tools that already talk to Ollama can switch to it with a one-line model name change.
Generate an API Key
Visit ollama.com/settings/keys while signed in, click "Create API key," and copy the value.
export OLLAMA_API_KEY=your_api_key_hereCall Muse Spark Open from the Local Endpoint
curl http://localhost:11434/api/chat -d '{
"model": "muse-spark-open:cloud",
"messages": [
{ "role": "user", "content": "Refactor this function to handle null inputs safely and add type hints." }
],
"stream": false
}'Python Example
from ollama import Client
client = Client(host="https://ollama.com", headers={"Authorization": "Bearer " + api_key})
response = client.chat(
model="muse-spark-open:cloud",
messages=[{"role": "user", "content": "Plan a database migration from PostgreSQL 14 to 17 with zero downtime."}],
)
print(response["message"]["content"])Launch Directly Into an Agent Scaffold
Ollama ships launch shortcuts that point an agent framework at a given model in one command, without hand-editing a config file first:
ollama launch claude --model muse-spark-open:cloud
ollama launch opencode --model muse-spark-open:cloud
ollama launch hermes --model muse-spark-open:cloud
ollama launch openclaw --model muse-spark-open:cloudEach of these drops you straight into that agent's interface running on Muse Spark Open, which is a faster path than manually editing Hermes Agent or OpenClaw's config file to point at `http://localhost:11434/v1` with the model name set.
Troubleshooting
`ollama pull muse-spark-open` fails with "pull model manifest: file does not exist"
Cause: Ollama's library only lists `muse-spark-open:cloud`, with no local or quantized tag through Ollama itself
Fix: Use `ollama run muse-spark-open:cloud` with the `:cloud` suffix included. A local tag may appear later, but it is not available through Ollama as of this guide's publish date.
`ollama run muse-spark-open:cloud` returns "model not found"
Cause: The installed Ollama version predates cloud model support
Fix: Update Ollama by re-running the install command, then retry. Cloud models require Ollama 0.6.x or later.
"unauthorized" error or repeated sign-in prompts
Cause: The machine is not signed in, or the session expired
Fix: Run `ollama signin` again and complete the browser approval. Check ollama.com/settings/connections to confirm the device is listed as connected.
First response takes 10-20 seconds or longer
Cause: Cold start while Ollama establishes a session with Meta's cloud infrastructure
Fix: This is normal for the first request after signing in or after an idle period. Subsequent requests in the same session stream back at normal speed.
API requests to `https://ollama.com/api` return 401
Cause: Missing or invalid `OLLAMA_API_KEY`
Fix: Generate a new key at ollama.com/settings/keys and re-export the environment variable.
Trying to run the Hugging Face weights directly runs out of memory
Cause: Loading all 816B parameters (even with only 42B active per token) requires hundreds of GB of combined VRAM at low quantization
Fix: Use a quantized build (Unsloth's GGUF releases are the common path) and a multi-GPU setup, or rent one on Vast.ai instead of buying hardware for a single model.
Alternatives to Consider
| Tool | Type | Price | Best For |
|---|---|---|---|
| Muse Glimmer | Local (Ollama) | Free, local install | Meta's 30B distillation of this same model, with real local Ollama tags starting at 18GB for a single GPU or Apple Silicon Mac. |
| Kimi K2.6 via Ollama Cloud | Cloud (Ollama) | Free within Ollama Cloud limits | Another cloud-only flagship with strong multi-agent and tool-use performance. |
| GLM 5.2 via Ollama Cloud | Cloud (Ollama) | Free within Ollama Cloud limits | A 744B cloud model with a 1M context window and effort-level control, also cloud-only through Ollama. |
| DeepSeek R1 | Local (Ollama) or VPS | Free | A true local reasoning model on hardware from 4GB (distilled) up to 64GB or more, with no cloud dependency. |
Frequently Asked Questions
Can I run Muse Spark Open locally with Ollama?
Not through Ollama's pull mechanism. Ollama currently lists only `muse-spark-open:cloud`, with no local tag. The model's weights themselves are public on Hugging Face under the Muse Model License, and you can run them locally through vLLM, SGLang, TensorRT-LLM, or llama.cpp if you have the hardware, but Ollama has not added a downloadable tag for this specific model.
If you want a version of this model you can actually pull and run through Ollama on a single GPU or Mac, see Muse Glimmer in the alternatives section, Meta's 30B distillation of this same base model.
Is Muse Spark Open free to use through Ollama?
Through Ollama Cloud, usage is billed at $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens as of this guide's publish date. `ollama signin` itself requires no payment information, but running prompts against `muse-spark-open:cloud` draws against Ollama's cloud usage billing once any free allocation is exhausted.
Check ollama.com/settings for your current usage and billing status, since allocations and pricing can change over time.
What is the difference between Muse Spark Open and Muse Glimmer?
Scale. Muse Spark Open is the full 816B-parameter mixture-of-experts model (42B active), Meta's largest open-weight release, while Muse Glimmer takes that same base and distills it down to 30 billion parameters, small enough to run locally on a single consumer GPU or an Apple Silicon Mac with enough unified memory.
If you have the hardware for an 816B model (or want to use Ollama's cloud route), Muse Spark Open is the more capable base. If you want something you can actually download and run today on normal hardware, Muse Glimmer is the practical choice.
Does Muse Spark Open support image input?
Yes. Muse Spark Open accepts text, image, audio, and video input and produces text output. Through Ollama, drop a file path into your prompt and Ollama attaches it automatically, the same mechanism used by any multimodal model in Ollama's library.
Can I use Muse Spark Open commercially?
It depends on your usage, since Muse Spark Open ships under the Muse Model License, which Meta describes as having limited commercial terms rather than a fully permissive license like Apache 2.0 or MIT. Read the license text on the model's Hugging Face page before deploying it in a commercial product.
Muse Glimmer, by contrast, ships under Apache 2.0, a fully permissive license with no comparable restriction.
What hardware does Muse Spark Open need to run locally, outside Ollama?
Meta's own reference setup for running the full model locally is two Mac Studios with M5 Ultra chips. Outside that specific configuration, expect to need a multi-GPU server with hundreds of gigabytes of combined VRAM, even with quantized weights, since 816B total parameters is a genuinely large model regardless of how many are active per token.
Renting a multi-GPU node on a service like Vast.ai by the hour is a more realistic option than buying hardware for a single model.
How does Muse Spark Open compare to other flagship models on benchmarks?
Meta has not published formal benchmark scores for Muse Spark Open as of this guide's publish date. Treat any third-party comparison to models like GLM 5.3 or Kimi K3 as unverified until independent evaluations are published and confirmed.
Can I use Muse Spark Open with an agent like Hermes Agent or OpenClaw?
Yes, and Ollama provides a shortcut for it. Running `ollama launch hermes --model muse-spark-open:cloud` or `ollama launch openclaw --model muse-spark-open:cloud` drops you directly into that agent's interface running on Muse Spark Open, without manually editing a config file first.
You can also configure Hermes Agent or OpenClaw manually by pointing their model configuration at Ollama's OpenAI-compatible endpoint, `http://localhost:11434/v1`, with the model name set to `muse-spark-open:cloud`.
Do I need a VPS or GPU to use Muse Spark Open through Ollama?
No. Inference for the `:cloud` tag runs on Meta's infrastructure regardless of where you run `ollama`, so a laptop or desktop with no GPU is enough for the cloud route covered in this guide. A GPU or multi-GPU server is only needed if you run the published weights yourself outside Ollama.
Related Guides
How to Run Muse Glimmer Locally with Ollama (2026 Guide)
How to Run MiniMax M3.1 Flash on Ollama: Cloud Setup Guide (2026)
How to Run Kimi K2 on Ollama: Cloud Setup Guide (2026)
How to Run GLM 5.2 on Ollama: Cloud Setup Guide (2026)
How to Run DeepSeek R1 Locally with Ollama (2026 Guide)
Best Local LLM Models to Run in 2026 (Benchmarks + Use Cases)
How to Install Hermes Agent with Ollama Local Models (2026)
How to Run OpenClaw with Ollama Local Models (2026 Guide)