Can you download Muse Wonder weights?
The official catalog documents Wonder through Model API and Muse Code, and identifies Spark Open and Glimmer as the downloadable model routes. We did not verify an official Wonder weight download or a Wonder open-source license. Do not substitute a similarly named checkpoint or a third-party file for an official release.
This is an evidence statement about the current catalog, not a prediction about future releases. Check the provider’s model documentation before planning a local Wonder deployment. If your requirement is specifically self-hosting, evaluate the releases that actually include weights.
Two documented self-hosted options
| Release | Official description | Deployment consideration |
|---|---|---|
| Muse Spark Open | 816B total / 42B active, mixture of experts | Server-class weight and runtime memory |
| Muse Glimmer | 30B dense, text and image input | A smaller local/edge-oriented release |
Spark Open’s official overview describes BF16 and NVFP4 checkpoints and a Muse Model License. Glimmer’s overview identifies Apache 2.0 licensing. These are different licenses: “open-weight” is not a promise that every Muse model shares one set of permissions.
Weights are not a free inference service
Build a self-hosted budget from GPU or machine rental, memory, storage, data transfer, runtime operations and expected utilization. An idle server still costs money. A hosted API charges on a different basis, so compare the same completed workload at realistic utilization.
Before buying hardware, read the checkpoint size and the serving runtime’s memory needs. Weight size alone excludes runtime overhead, activations and the key-value cache. Quantization changes the deployment constraints and may change behavior; it should be recorded in your evaluation.
A hypothetical $2/hour server used for 100 hours costs $200 before other fees. That is arithmetic with a chosen assumption, not a quote for hardware capable of serving either release. Replace it with an actual provider quote and a measured throughput before estimating cost per accepted task.
Verify a download and rollout
- Follow the download links from the official model documentation.
- Read the model license and acceptable-use conditions for the selected checkpoint.
- Confirm the runtime supports its architecture, quantization and intended context.
- Test a small reproducible workload and record the exact checkpoint revision.
- Benchmark utilization and cost before committing to a long hardware rental.
For a hosted route, use the API onboarding guide. For a free consumer entry point, see free app access versus paid developer usage. Neither route proves that Wonder weights are downloadable.
Sources & evidence
Editorial check: 17 October 2026. Product facts come from the following primary sources; planning checklists and examples are our explanations. No paid inference, model benchmark or regional checkout was run for this guide.