Meta Muse Spark Open (816B-A42B) open-weights model card

What it claims

Muse Spark Open is an 816B total / 42B active open-weights checkpoint. Meta says it has extra safety post-training so it holds up as a standalone model without system guardrails. Meta still recommends deploying it inside a guarded system. Inputs include images, audio and video, and the card covers screenshots, charts, documents and recordings in agentic settings.

Evidence

Meta reports cyber-misuse refusal rates measured with adaptive attackers who query the model like an API, with no system-level safeguards:

The card says these results apply only to the released checkpoint. Modified or fine-tuned models should not be assumed to keep them.

How to use it

Access is gated. You must log in to Hugging Face and submit your full legal name, date of birth and full organization name. The form cannot be edited after submission, and errors "may prevent you from accessing this model and others on Hugging Face."

Hardware. Two Mac Studios with M5 Ultra under MLX are named as an optimized target.

Serving:

vllm serve "meta-models/Muse-Spark-Open-816B-A42B"

python3 -m sglang.launch_server \
    --model-path "meta-models/Muse-Spark-Open-816B-A42B" \
    --host 0.0.0.0 \
    --port 30000

docker model run hf.co/meta-models/Muse-Spark-Open-816B-A42B

Transformers. Use the image-text-to-text pipeline, or AutoProcessor with AutoModelForMultimodalLM. Both servers expose OpenAI-compatible /v1/chat/completions.

Inference settings:

• reasoning_effort accepts only low, medium, high or xhigh. Any other value returns an error.
• You can pass it as a request parameter or via chat_template_kwargs.
• The default is xhigh. Set low or medium explicitly if you want short, fast responses.
• Recommended sampling: temperature 1.0, top_p 1.0, top_k disabled.

Licence obligations:

flowchart TD
  A[Use Software or Derivative] --> B[Show 'Built with Muse Spark Open']
  A --> C{Offered as Model-as-a-Service?}
  C -- No: internal, research, own product feature --> D[No extra agreement]
  C -- Yes --> E{Group revenue > $20M in any 12 months?}
  E -- No --> D
  E -- Yes --> F[Notify Meta and sign separate agreement]

• "Derivatives" include any model trained, fine-tuned or distilled using the model's outputs. A model trained on synthetic data from Muse Spark Open therefore inherits the attribution requirement and the Usage Policy.
• The $20M revenue test counts the licensee plus affiliates.
• The MaaS clause does not apply when the model powers features of your own non-MaaS product.
• Meta claims no ownership of outputs.

Limits

Usage Policy restrictions:

• Prohibits use for military, warfare, nuclear, espionage or ITAR-controlled work.
• Prohibits operation of critical infrastructure, transportation or heavy machinery.
• Prohibits unlicensed professional practice (financial, legal, medical).
• Prohibits representing outputs as human-generated.
• Prohibits circumventing safety measures.
• The model is not for users under 18. Deployers must assess and mitigate risk where minors may reach it.

Stated failure modes:

• Very short or low-resolution video (a few seconds at about 100 px) can make the model say it can't view the video. Use 448 px or more on the short side.
• Multi-step reasoning errors in novel scenarios.
• Degraded quality on languages outside an unnamed "strongly supported set."
• Minor edge-case differences when running quantized.

Deployment guidance:

• Do not expose raw chain-of-thought.
• Protect internal assistant messages from user manipulation.
• Add human-in-the-loop confirmation for irreversible agentic actions.

Missing benchmark comparison. The card says it compares against open-weight models of similar size, but that comparison table is not in the fetched text. The listed scores (GPQA Diamond, Terminal-Bench 2.1, Toolathlon and others) have no baselines, and several carry unexplained asterisks. Check the linked report before relying on capability claims.