Muse Spark Open: Specs, License & How to Run

Muse Spark Open is a gated open-weight multimodal model from Meta. Official hosts we opened on 2026-10-16: the gated card meta-models/Muse-Spark-Open-816B-A42B, the NVFP4 sibling meta-models/Muse-Spark-Open-816B-A42B-NVFP4, the collection meta-models/muse-spark-open, the Muse Model License, the launch post Introducing Muse Spark Open (15 October 2026), and dev.meta.ai docs.

This identification page hangs off ChatGPT alternatives and closed LLM APIs. Those mappings are editorial. We have no sourced quality comparison against ChatGPT.

Do not confuse this Meta card with Spark-X2.6-4B (XHToken / SparkLLM Team, ungated Apache-2.0 ~4B on-device text). Same word, different project. On-device peer compare: Spark-X2.6-4B vs LFM2.5-24B-A2B.

The Meta blog (15 October 2026) calls it a natively multimodal model with 816B total / 42B active parameters and a 1M-token context, and says weights ship under the Muse Model License with limited commercial terms. Hugging Face pipeline_tag is image-text-to-text. The blog’s “most capable US open-weight model” sentence is a vendor claim, not a forkalt ranking.

What we verified (2026-10-16)

Field	Value	Source
HF likes / downloads	264 / 402	HF API meta-models/Muse-Spark-Open-816B-A42B
Card created / last modified	2026-09-15T19:23:19Z / 2026-10-15T15:46:43Z	same
Gated	manual	HF API
License (card)	other / muse-model-license	HF cardData
LICENSE file	Muse Model License, version Muse Spark Open, 15 October 2026	LICENSE.md
Pipeline tag	image-text-to-text	HF API
Params (Meta)	816B total / 42B active sparse MoE	Blog + docs + HF card
Context	1,048,576 tokens	Meta docs + HF card
Modalities	text + image + audio + video in → text out	Meta docs + HF card
Architecture (card)	256 experts, top-8; 62 LM layers; PE 1.8B / 50 layers	HF card “Model Overview”
NVFP4 sibling likes / downloads	104 / 331	HF API …-NVFP4
NVFP4 gated	manual	same
NVFP4 usedStorage	533,039,109,326 bytes (~496 GiB on disk; Meta docs say ~537 GB weights)	HF API + Meta docs

The BF16 card’s usedStorage in the same API check is 1,644,361,479,718 bytes (~1.53 TiB / ~1.65 TB), which matches the Meta docs’ BF16 weight-size line. Hugging Face safetensors.parameters on the BF16 card sum to 822,036,529,664 (~822B). We keep Meta’s 816B / 42B quotes and do not invent a reconciliation.

Hardware (vendor / datacenter-class)

runsLocally: true here means the weights are downloadable after the gate, not that a consumer GPU will load them.

Copied from Meta docs and the HF card (checked 2026-10-16):

Checkpoint	Weight size (Meta)	Serving note (Meta)
NVFP4	~537 GB	≥700 GB VRAM (card); 4-GPU NVIDIA Blackwell node (docs)
BF16	~1.65 TB	~2 TB VRAM budget (docs); ~8 × B300 (docs)
Apple	—	two Apple M5 Ultra Mac Studios via MLX (blog + card)

Runtimes Meta names: SGLang, vLLM, TensorRT-LLM, Transformers, Apple MLX, llama.cpp, Unsloth. Hosted partners Meta names: Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together. Catalogs named on the blog: Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Models, Hugging Face. Those are copied names. forkalt has not opened each partner’s model SKU page.

How to run (from official docs, not yet tested)

The card and dev.meta.ai/docs/muse-spark-open point at self-hosting with SGLang, vLLM, or another named runtime, or calling a hosted provider. Recommended sampling copied from the card: temperature 1.0, top_p 1.0, top_k disabled; reasoning effort high or xhigh for complex / agentic work (default xhigh). forkalt has not run these commands.

License

LICENSE.md is titled MUSE MODEL LICENSE, version Muse Spark Open, release date October 15, 2026, copyright Meta Platforms, Inc. Hugging Face cardData.license is the opaque label other. We do not invent a more specific SPDX id.

Paraphrase only, not legal advice:

• Include the copyright notice, this license, and the Usage Policy with copies / substantial portions and Derivatives.
• Prominently display “Built with Muse Spark Open” on products or services that use the Software or Derivatives (or in accompanying docs / website / settings).
• Model-as-a-Service threshold: if you give third parties model access via API or a similar interface and aggregate gross revenue of you and affiliates exceeds USD 20 million in any consecutive 12 months, notify Meta and enter a separate written agreement before offering the Software or Derivatives as part of that MaaS business. Section 3 does not apply to (a) internal use, research, or commercial activity that is not MaaS, or (b) using the Software as a component of your own product that is not MaaS.
• Meta claims no ownership of outputs.
• Usage Policy (Exhibit A, 15 October 2026): the model is not intended for individuals under 18. The policy lists prohibited uses (law, harm, deception, and related categories). Read the file.

What we did not verify

Vendor benchmark tables on the card (JobBench, MCP-Atlas, GPQA Diamond, Terminal-Bench, and others versus named open-weight models) stay self-reported. Safety refusal percentages on the card are upstream. We have no sourced bake-off against ChatGPT. We have not loaded NVFP4 or BF16 on a named cluster.

Related chat weights on this site: Mistral Large 4 (Mistral Large 4 vs Muse Spark Open), Spark-X2.6-4B (different “Spark”; Apache-2.0 ~4B on-device), Reflection Beam (Beam vs Muse Spark Open), DeepSeek-V4.1-Flash, Kimi K3, and MiniCPM-V 4.7 on ChatGPT alternatives; Kolibri-1 and LFM2.5-24B-A2B on closed LLM APIs. Runtimes: vLLM, DeepInfer vs vLLM. Catalog: home. Editorial: About. Pixel editors: PhotoCraft vs GIMP.