Learn about this first
On October 15, Meta released Muse Spark Open, an open-weight model using a MoE architecture, with 816B total parameters and 42B active per token. It has a 1M-token context window, natively supports text, image, audio, and video input, and is designed for coding and agent tasks. Meta calls it the strongest US open-weight model to date; Artificial Analysis gave it an Intelligence Index score of 39, ranking it first among US open-weight models.
The weights are available on Hugging Face under the Muse Model License in BF16 and NVFP4 formats. The license requires attribution with “Built with Muse Spark Open.” Model-as-a-Service providers with annual revenue exceeding $20 million must sign a separate agreement with Meta; the revenue clause does not apply to internal use, research, or features in their own products. On launch day, platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI made APIs available. Together AI and OpenRouter charge $1.25 per million input tokens and $4.25 per million output tokens, with cached input priced at $0.15.
Apple and Meta worked together to adapt MLX, demonstrating local execution on two M5 Ultra Mac Studio systems. vLLM and SGLang offer Day-0 support, although vLLM’s PyPI wheel will take about a week to appear in the nightly version after the PR is merged. Third-party benchmark results vary: Arena reports that Muse Spark Open (Max) scored 1511 in Code Arena: WebDev, ranking first among U.S. open-source models and 13th overall among open-source models. ARC Prize reports scores of 89.0% and 57.5% on ARC-AGI-1/2, comparable to Haiku 5.5 and GPT-6 Luna, but at about 9 times the latter’s cost; its ARC-AGI-3 score is just 0.4%. Meta Chief AI Officer Tao Wang said the company generally does not open-source its most advanced models, but will continue releasing strong open-source models.
AI-generated from reports · updated 19 days agoLatest developments
Meta Muse Spark ARC-AGI Benchmark Results Announced
Event progress14
ARC Prize Announces Muse Spark’s ARC-AGI Scores
ARC Prize: Meta Muse Spark ARC-AGI Benchmark Results AnnouncedArtificial Analysis publishes performance test results for the Muse Spark Open model
Artificial Analysis: Artificial Analysis publishes measured inference-throughput results for Meta Muse Spark Open on AA-AgentPerfFireworks launches Muse Spark Open
Fireworks AI: Meta's open-weight model Muse Spark Open launches on Fireworks serverless in the US regionArena announces Muse Spark Open ranks first on the Code Arena WebDev leaderboard
Arena: Muse Spark Open (Max) tops Code Arena: WebDev as the No. 1 US open-source modelArena adds Muse Spark Open to Agent Arena
Arena: Muse Spark Open is now on Arena and available to test in Agent Arena; scores coming soonMuse Spark Open is now on OpenRouter
OpenRouter: Meta's open-weight model Muse Spark Open is now on OpenRouterMeta's Muse Spark Open earns praise
David Singleton: Meta's Muse Spark Open earns praiseOptimizing the runtime environment for Muse Spark Open
AI at Meta: Meta outlines ways to run Muse Spark Open on Mac Studio and major inference enginesMLX distributed LLM inference tutorial
Apple Machine Learning Research: Apple demonstrates running an 816B model locally on a Mac Studio cluster using MLX and RDMAvLLM announces Day 0 support for Muse Spark Open
vLLM Official Blog: vLLM provides Day 0 support for Muse Spark OpenTogether AI launches Meta's Muse Spark Open model
Together AI Research and Product Blog: Meta's Muse Spark Open is now on Together AIMuse Spark Open is now on Modal Auto Endpoints
Modal Official Engineering Blog: Meta open-sources Muse Spark Open; Modal launches shared endpoints with per-token billingSGLang and Miles provide Day-0 support for Muse Spark Open
LMSYS Blog: SGLang provides Day-0 support for Muse Spark Open, an 816B open-source multimodal MoE modelMuse Spark Open weights released
PyTorch: Blog: Meta releases open-weight model Muse Spark Open with 816B parameters and a 1M context window