Overview
Meta Muse Spark is Meta's next-generation AI large model released on 2026-04-08, and the first public achievement since the establishment of Meta Superintelligence Labs (MSL). Led by former Scale AI CEO Alexandr Wang, MSL built this model from scratch over 9 months.
What draws the most attention to Muse Spark is not the technology itself, but the strategic signal behind it—Meta abandoned the open-source route and shifted to closed-source. This marks a major strategic shift in Meta's AI model strategy following the Llama series. It ranks among the top five globally in comprehensive capability, is natively multimodal, supports three reasoning modes (Fast, Balanced, Deep), and is currently freely accessible via meta.ai.
📌 Additional Information: As of July 2026, Muse Spark remains the flagship model of Meta Superintelligence Labs (MSL) and serves as the native large model foundation for Meta Glasses 2026 (released on June 23). Meta's strategy is to elevate Muse Spark from a "model" to a "native capability for AI hardware entry points"—no longer selling APIs separately, but distributing it through hardware products like smart glasses and AI assistants. This also allows Meta to regain a foothold in the AI race.
The model has continued to advance through 2026. On 2026-09-02 Meta shipped Muse Spark 1.3, with notable gains in agentic and scientific reasoning: more efficient tool calls, lower token usage, and a 1M-token context window; third-party Intelligence Index scores land around 61-62, close to flagship Claude and GPT-5.6-tier models of the same period. On 2026-09-22, after the latest Artificial Analysis update, Muse Spark sits ahead of xAI Grok 4.7 on the composite intelligence index, even though Grok 4.7 launched the same day. On 2026-10-02, Meta published six papers completed by mathematicians working with Muse Spark 1.1 and 1.2 (Thinking Mode) through the ordinary meta.ai chat interface, with no custom research scaffolding; five of them answer previously open research questions across probability, differential equations, group theory, optimization, arithmetic physics and non-associative algebra, and each paper marks which sections were drafted by humans or AI, credits the prior work it depends on, and was reviewed by a second group of mathematicians.
Key Features
- Native Multimodality: Unified understanding and generation of text, images, and video
- Three Reasoning Modes: Fast Response / Balanced Mode / Deep Thinking, switchable on demand
- Top Five Globally: Comprehensive intelligence index ranks among the world's top models
- Meta Ecosystem Integration: Full connectivity across WhatsApp, Instagram, and Facebook
- Completely Free: Free to use via the meta.ai web interface and Meta App
- MSL Debut: The first achievement from Meta's comprehensive rebuild of its AI R&D system
- Muse Spark 1.3: Released 2026-09-02 with gains in agentic and scientific reasoning, more efficient tool calls, lower token usage, and a 1M-token context window
- Dual-Track API Pricing: Standard Tier (muse-spark-1.3) input priced at $1.25/m
- Research-Grade Output: Six papers completed by mathematicians with Muse Spark 1.1 and 1.2 (Thinking Mode) via the ordinary meta.ai chat interface, published 2026-10-02
Use Cases
- Individual users needing free access to top-tier AI capabilities
- Heavy users of WhatsApp / Instagram / Facebook
- Multimodal content understanding and creation scenarios
- Early adopters wanting to experience Meta's latest AI capabilities
- Technical evaluations comparing different large model selections
- Agentic and scientific reasoning tasks, including long-context work with the 1M-token window
- Mathematical research and open-question exploration through the standard chat interface
Pros
- Completely free with very low barriers to entry
- Top five globally in comprehensive capability with solid performance
- Native multimodality, not a post-hoc patchwork
- Flexible switching between three reasoning modes
- Deep integration with Meta's social ecosystem
- Muse Spark 1.3 brings stronger agentic and scientific reasoning with more efficient tool calls and lower token usage
- 1M-token context window
- Third-party Intelligence Index scores around 61-62, close to flagship Claude and GPT-5.6-tier models of the same period
- Sits ahead of xAI Grok 4.7 on the Artificial Analysis composite intelligence index as of 2026-09-22
- Dual-track API pricing with a Standard Tier (muse-spark-1.3) at $1.25/m input
- Research-grade output: six mathematician co-authored papers completed via the ordinary meta.ai chat interface, five answering previously open research questions
Pricing
Completely free to use via meta.ai and Meta apps (WhatsApp, Instagram, Facebook). Meta has also introduced dual-track API pricing: the Standard Tier (muse-spark-1.3) is priced at $1.25/m input, a structure analyzed by Tom Tunguz on 2026-09-03 in terms of the data-to-compute logic behind it.
Summary
Muse Spark is a signal flare for Meta's restart after the Llama setback. The model itself is solid in capability and free to use, but the shift to closed-source subtly changes Meta's position in the developer ecosystem. It's good news for ordinary users (free access to top-tier models) but bad news for developers (no longer open-source). Through 2026 the line has kept moving: Muse Spark 1.3 arrived on 2026-09-02 with stronger agentic and scientific reasoning, a 1M-token context window, and third-party Intelligence Index scores around 61-62, and by 2026-09-22 it sits ahead of xAI Grok 4.7 on the Artificial Analysis composite intelligence index.
Version History
- A medical research scaffold pairing Muse Spark with Muse Wonder produces ten manuscripts in four weeks (2026-10-14): On October 14, 2026, Meta published progress in health data science. The team built an internal coding scaffold powered by a combination of Muse Spark 1.3 and Muse Wonder, with access to public data and a licensed retrospective database of five million de-identified electronic health records. A cohort of medical researchers with backgrounds in cardiology, epidemiology, infectious disease and internal medicine used the scaffold to conduct original research over four weeks, brainstorming hypotheses, writing and executing all code, and drafting manuscripts together with the model. The output is ten research manuscripts spanning arrhythmias after infective endocarditis, remission of type 2 diabetes, pickleball injuries, e-scooter and e-bike injuries, dental disease in S. aureus bacteremia, breast cancer drug sequencing and more. Each went through peer feedback, biostatistician review of the analysis, and summary review by contracted independent experts. Meta frames the system as human-led, with domain experts responsible for the scientific decisions and conclusions.
- Six papers co-authored with mathematicians (2026-10-02): Meta published six papers completed by mathematicians working with Muse Spark 1.1 and 1.2 (Thinking Mode) through the ordinary meta.ai chat interface, with no custom research scaffolding. Five of them answer previously open research questions across probability, differential equations, group theory, optimization, arithmetic physics and non-associative algebra. Each paper marks which sections were drafted by humans or AI, credits the prior work it depends on, and was reviewed by a second group of mathematicians
- Muse Spark moves ahead of Grok 4.7 on the Artificial Analysis leaderboard (2026-09-22): After the latest Artificial Analysis update, Meta Muse Spark sits ahead of xAI Grok 4.7 on the composite intelligence index, even though Grok 4.7 launched the same day. The two models were lined up against each other repeatedly the same week, and the shift in position renewed discussion of Muse Spark dual-track pricing.
- Tom Tunguz breaks down the data-to-compute logic behind Meta Muse Spark's dual-track pricing (2026-09-03): Tom Tunguz analyzes Meta's release of the Muse Spark model and its dual-track API pricing: Standard Tier (muse-spark-1.3) input $1.25/m.
- Meta ships Muse Spark 1.3 (2026-09-02): Meta released Muse Spark 1.3 with notable gains in agentic and scientific reasoning: more efficient tool calls, lower token usage, and a 1M-token context window; third-party Intelligence Index scores land around 61-62, close to flagship Claude and GPT-5.6-tier models of the same period.