Muse Spark

weights

open(BF16 and NVFP4 safetensors on Hugging Face, released after a manually approved access request)

data

closed(training data described only by source type: public, crawled, licensed, Meta products and synthetic

code

partial(chat template and configs, with inference through SGLang, vLLM, Transformers and other frameworks

license

Muse-Model-License(MIT-style grant

Meta publishes the Muse Spark Open weights for download after an access request, under a license that lets almost anyone use them commercially: only model-as-a-service businesses with more than USD 20 million in yearly revenue need a separate agreement with Meta. The training data and the post-training pipeline are not released. Muse Spark 1.3, the hosted model in Meta's apps and API, shares its base model and ships no weights.

• https://developer.meta.com/ai/models/muse-spark/ recorded 2026-10-19

  Meta developer site lists "Muse Spark 1.3" and "Muse Spark Open" as separate models; the Muse Spark 1.3 page offers it through the Meta Model API: "Muse Spark 1.3 is trained for agentic workflows and optimized for competitive coding performance".

• https://huggingface.co/api/models/meta-models/Muse-Spark-Open-816B-A42B recorded 2026-10-19

  Hugging Face API for meta-models/Muse-Spark-Open-816B-A42B: "gated": "manual", 65 model-*.safetensors files, cardData license_name "muse-model-license", and an extra_gated_prompt carrying the "MUSE MODEL LICENSE, version: Muse Spark Open, Release date: October 15, 2026" text: rights to "use, copy, modify (including fine-tune), ... distribute, sublicense, and/or sell", a "Built with Muse Spark Open" display requirement, the Usage Policy "hereby incorporated into this license", and Section 3: a Model as a Service licensee whose revenue "exceeds twenty (20) million US dollars ... during any consecutive twelve (12) month period" must "enter into a separate written agreement with Meta"; Section 3 "does not apply to (a) internal use, research, or any commercial activity that is not Model as a Service".

• https://huggingface.co/api/models/meta-models/Muse-Spark-Open-816B-A42B-NVFP4 recorded 2026-10-19

  Hugging Face API for meta-models/Muse-Spark-Open-816B-A42B-NVFP4: "gated": "manual", 127 safetensors files, cardData license_name "muse-model-license".

• https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B recorded 2026-10-19

  Muse Spark Open model card: "an 816B total/42B active parameter open weights model"; "The model is available in BF16 and NVFP4 versions"; supported frameworks "SGLang, vLLM, Hugging Face, TensorRT-LLM, Unsloth and Llama.cpp"; "Training data includes text, images, audio, and video. It includes data drawn from public sources, crawled from the Internet, acquired from third parties, derived from Meta's products and services, and synthetically generated". No training data or training code is linked.

• https://pytorch.org/blog/muse-spark-open/ recorded 2026-10-19

  PyTorch blog post by Meta: "Building on our August release of Muse Spark, this open weights version has additional safety post-training"; Muse Spark Open "shares the same base model as Muse Spark 1.3"; "available today on Hugging Face under the Muse Model License".

Muse Spark powers Meta AI in the Meta AI app and website and across WhatsApp, Instagram, Facebook, Messenger, Threads and Meta's AI glasses, and since October it also ships as the downloadable Muse Spark Open weights. Meta describes the reach of those surfaces as a whole and publishes no active-user figure for Muse Spark itself. Hugging Face downloads measure only the Open weights, one SKU of the line, so they do not band the product.

• https://about.fb.com/news/2026/04/introducing-muse-spark-meta-superintelligence-labs/ recorded 2026-10-20

  Meta newsroom post, updated through May 12, 2026: "Muse Spark now powers the Meta AI assistant in the Meta AI app and meta.ai"; "Muse Spark is starting to bring the same intelligence to Meta AI across WhatsApp, Instagram, Facebook, Messenger, and Threads"; "Muse Spark and the new Meta AI are reaching billions of people across our apps". No Muse Spark-specific user count is published.

Muse Spark 1.3 (max) ranks 9th on the arena.ai text leaderboard, about 15 points behind the top entry, and scores 48 on the Artificial Analysis Intelligence Index against a median of 26. Meta's own benchmark tables are not included here.

• https://arena.ai/leaderboard/text recorded 2026-09-29

  arena.ai text leaderboard, Sep 25, 2026: "View overall rankings across various AI models in text-to-text tasks"; "muse-spark-1.3-max" (Meta · Proprietary) at rank 9, score 1494 +/-7 over 10,036 votes.

• https://artificialanalysis.ai/models/muse-spark-1-3 recorded 2026-09-29

  Artificial Analysis page for Muse Spark 1.3 (max): "scores 48 on the Artificial Analysis Intelligence Index, placing it well above average among other reasoning models in a similar price tier (median: 26)".