Meta · Muse

Muse Spark

Muse Spark is Meta Superintelligence Labs’ lighter Muse model: a natively multimodal reasoning model for agentic work, coding, and Meta AI. Current developer pages highlight Muse Spark 1.3. It is available through Meta AI products and the Meta Model API (public preview language on Meta pages). Since 12 Oct 2026 Meta’s most capable model to date is Muse Wonder, which builds on Spark.

Status: Current, Meta's lighter Muse model

Training

Meta frames Spark as a rebuilt pre-training stack (architecture, optimization, data). Blog posts claim Muse Spark reaches comparable capability with over an order of magnitude less training compute than Llama 4 Maverick. Exact parameter count, FLOPs table, and dataset mix for Spark 1.3 are not published on the pages cited here. Treat the compute-efficiency claim as a vendor claim.

Inference

Multimodal perception (video, images, documents on Meta’s Spark product pages). Tool use and agentic workflows. Reasoning modes appear on Meta blogs (for example max / Contemplating language). The Meta Model API is documented in an OpenAI-compatible style (api.meta.ai).

Context

Developer Muse Spark materials list a 1M token context for muse-spark-1.3 SKUs. Confirm on the live developer model page.

Methods

Multimodal reasoning, tool use, visual chain of thought, and multi-agent orchestration are named on the MSL introduce blog. Exact train-time RL recipe is not published on those pages.

Benchmarking

Meta publishes benchmark tables on the Muse Spark developer page and evaluation reports. Use only scores that appear on Meta’s live page or a signed eval PDF. Cross-lab comparisons on Meta’s table are vendor-reported.

Hardware

Meta has not published training or serving silicon details for Muse Spark on the model pages cited here. Hyperion and other infra claims belong on the hardware map when Meta publishes them.

Access and pricing

API ids named on Meta’s Muse Spark model page include muse-spark-1.3 and muse-spark-1.3-contributor. Dual price tiers and cache rates appear on that page. Prefer live prices at ship. The contributor tier is described as used to improve Meta products; the higher-priced tier as not used to improve products (Meta wording).