What Is Muse Spark?
Muse Spark is a natively multimodal reasoning model from Meta Superintelligence Labs (MSL), the research division Meta stood up in 2025 under Alexandr Wang. It reads text and images, returns text, and reasons in three modes — Instant, Thinking, and Contemplating. Meta released the first version on April 8, 2026 and shipped Muse Spark 1.1 on July 9, 2026.
Meta Superintelligence Labs is Meta's frontier-AI group, formed in 2025 and led by Chief AI Officer Alexandr Wang, the former CEO of Scale AI, with Nat Friedman and former OpenAI researcher Shengjia Zhao. Muse Spark is its flagship model — a departure from Meta's earlier open-weight Llama line, rebuilt from scratch as a proprietary reasoning system.
The headline feature is Contemplating mode. Instead of a single pass, Muse Spark can spawn several reasoning agents that work a problem in parallel and combine their answers, which Meta pits against the heavy reasoning modes of Gemini Deep Think and GPT Pro. Lighter Instant and Thinking modes cover everyday questions and standard chain-of-thought.
Muse Spark Technical Details
Muse Spark 1.1 runs a 1 million-token context window, up from 262K on the April release, so entire codebases, long contracts, and book-length documents fit in one conversation. It accepts text and image input and returns text, and Meta trained its health responses with input from more than 1,000 physicians.
Muse Spark is proprietary — unlike Meta's earlier Llama models it is not open-weight and cannot be self-hosted. Meta serves it through meta.ai, the Meta AI app, and the Meta Model API (public preview). Third-party platforms like Overchat AI put it alongside competing models such as GPT-5.6, Claude Opus 4.8, Gemini 3 Pro, and DeepSeek V3.2.
On benchmarks, Muse Spark leads on health: it scores 42.8 on HealthBench Hard, ahead of GPT-5.4 (40.1) and Gemini 3.1 Pro (20.6), and refuses bioweapon prompts 98.0% of the time. It lands mid-pack elsewhere — Artificial Analysis put the original fourth on its Intelligence Index behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6, and it trailed on Terminal-Bench 2.0 (59.0 vs GPT-5.4's 75.1). The 1.1 update closes much of the coding gap.
Muse Spark 1.1 leans hard into agentic coding: it gathers context, plans, and delegates work to parallel subagents to cut end-to-end latency on large, real-world codebases. Meta says it surpasses its own earlier model on internal coding evals, and Zuckerberg has promised aggressive API pricing against OpenAI, Anthropic, and Google.
.svg){kind=icon}
