Skip to content
Muse Spark Open logo

Muse Spark Open

Meta

Key facts · verified Oct 18, 2026

Muse Spark Open is a Meta model released on 15 Oct 2026. Pricing: Free - open weights, self-hosted (no first-party API price). Its context window is 1,048,576 tokens. It is not yet ranked on LMArena's text leaderboard. Source: dev.meta.ai.

Muse Spark Open is the open-weight release of Muse Spark, Meta's agentic and coding model. It is a sparse mixture-of-experts transformer with 816 billion parameters in total and 42 billion active per token, with built-in perception encoders: it takes text, image, audio and video input and returns text, with a 1,048,576-token context window. Meta publishes it in two checkpoints - BF16 (about 1.65 TB of weights, served on eight GPUs such as 8 x B300) and NVFP4 (about 537 GB, served on one 4-GPU Blackwell node) - under the Muse Model License. It can be self-hosted with runtimes such as SGLang and vLLM, or called through hosted providers.

Pricing & specs checkedSource: vendor documentation · checked Oct 18, 2026
Pricing

Free - open weights, self-hosted (no first-party API price)

Context

1,048,576 tokens

LMArena Elo

Not rated

LMArena rank

Not ranked

Full Specifications

Free - open weights, self-hosted (no first-party API price)
Identity
DeveloperMeta
ReleasedOct 2026
StatusGA
LicenceMuse Model License, plus Meta's Acceptable Use Policy
Self-hostableYes - BF16 (~1.65 TB, 8 x B300) or NVFP4 (~537 GB, 4 x B300); weights on Hugging Face (meta-models/Muse-Spark-Open-816B-A42B), gated
Cost
Input priceSelf-hosted - no first-party API price
Output priceSelf-hosted - no first-party API price
Free tierYes - weights are free to download and run
Capacity
Context window1,048,576 tokens
Capability
Vision inYes - image and video
Audio inYes
Function callingYes - tool calling (per Meta's prompting guide)
Extended reasoningYes
Access
APINot offered first-party (hosted providers and OpenRouter)
Chat appNo
Cloud marketplacesAmazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Model Catalog
Fine-tuningYes - open weights; Meta publishes a post-training guide