What is Muse Spark? Meta's New AI Model (April 2026)
What is Muse Spark? Meta’s New AI Model (April 2026)
Muse Spark is Meta’s first major AI model since the $14 billion Scale AI acquisition and Alexandr Wang’s move to lead Meta Superintelligence Labs. Launched April 8, 2026, it is Meta’s clearest attempt yet to compete at the frontier with OpenAI, Anthropic, and Google.
Last verified: April 19, 2026
The quick answer
Muse Spark is a natively multimodal Mixture-of-Experts (MoE) model with a 1M-token context window, released free in Meta’s consumer apps — meta.ai, WhatsApp, Instagram, and Messenger. It is the highest-scoring free frontier-class model on Artificial Analysis’s Intelligence Index (52), trailing GPT-5.4 (57), Gemini 3.1 Pro (57), and Claude Opus 4.6 (53).
Why it matters
Meta spent years being overshadowed in the LLM race. Llama 4 was well received among open-model users but never matched GPT-5 or Claude Opus at the frontier. The $14B Scale AI acquisition brought in Alexandr Wang to run Meta Superintelligence Labs, and Muse Spark is his first major output.
Three reasons Muse Spark matters:
- It’s genuinely competitive on multimodal perception — beating GPT-5.4 and Claude on MMMU-Pro (80.5% vs 78.3% vs 75.8%).
- It’s free. Zero-cost frontier-class AI at the scale of WhatsApp (3B users) reshapes the economics of the industry.
- It signals Meta’s strategy. Premium proprietary model for consumers (Muse Spark); Llama continues as the open-source research track.
Key facts (April 19, 2026)
| Property | Value |
|---|---|
| Released | April 8, 2026 |
| Developer | Meta Superintelligence Labs |
| Lead | Alexandr Wang (chief AI officer) |
| Architecture | MoE, natively multimodal |
| Context window | 1M tokens |
| Modalities in | Text, image, audio |
| Modalities out | Text, image |
| Access | meta.ai, WhatsApp, Instagram, Messenger, Ray-Ban Meta glasses |
| API | Partner waitlist (no public pricing yet) |
| Open weights | Not yet — planned “in the coming months” |
Benchmarks
| Benchmark | Muse Spark | GPT-5.4 | Claude Opus 4.7 | Gemini 3.1 Pro |
|---|---|---|---|---|
| AA Intelligence Index | 52 | 57 | 55 | 57 |
| HLE | 39.9% | 41.6% | 40.7% | 44.7% |
| MMMU-Pro (vision) | 80.5% | 78.3% | 75.8% | 82.4% |
| CharXiv (chart reasoning) | leader | — | — | — |
| Terminal-Bench 2.0 (coding) | 59.0 | 75.1 | 78.4 | 68.5 |
| GDPval-AA (agent Elo) | 1,444 | 1,672 | 1,606 | — |
| ARC-AGI-2 | 42.5 | ~74 | ~72 | ~75 |
Strengths: Multimodal perception, chart/data reasoning, general chat, cost (free).
Weaknesses: Coding, long-horizon agent tasks, abstract visual reasoning (ARC-AGI-2).
What you can do with it
Since Muse Spark is embedded in Meta’s apps, most use cases live in consumer surfaces:
- WhatsApp / Messenger chat — voice and text, with memory
- Instagram — captions, edits, visual suggestions
- meta.ai — full chat, image generation, search
- Ray-Ban Meta glasses — real-time visual assistance (“what’s this?”)
- Image editing — native output, including layout changes
How to access Muse Spark
| Channel | How |
|---|---|
| meta.ai (web) | Visit meta.ai, sign in with Meta account |
| Chat with the Meta AI contact | |
| @ tag Meta AI in DMs or search | |
| Messenger | Direct chat with Meta AI |
| Ray-Ban Meta | ”Hey Meta, …” voice prompt |
| API | Join partner waitlist on ai.meta.com |
It is completely free at point of use in consumer apps as of April 19, 2026. No subscription, no token cap advertised to users.
How it differs from Llama 4
| Dimension | Llama 4 | Muse Spark |
|---|---|---|
| License | Llama Community (700M MAU cap) | Closed, open weights TBD |
| Target users | Developers / researchers | Consumers (via Meta apps) |
| Architecture | Dense + some MoE variants | Full MoE, natively multimodal |
| Distribution | Hugging Face, Ollama, direct | Meta’s consumer apps |
| Positioning | Open-source flagship | Proprietary frontier flagship |
Muse Spark is essentially Meta’s answer to GPT and Claude, while Llama continues as the open-source track. Expect both lines to coexist: Llama 5 (when it arrives) will remain the research/open release; Muse Spark family will be Meta’s commercial top end.
What comes next
Meta’s blog post mentioned “larger models in development” — implying Muse Spark is only the start of a multi-model family. Reasonable bets:
- A larger Muse Spark Ultra at some point in 2026
- A coding-specialized variant (to close the Terminal-Bench gap)
- Open-weights release of the current Muse Spark later in 2026
- Deeper integration into Ray-Ban Meta and Meta Quest devices
Bottom line
Muse Spark is the biggest “free AI” story of 2026. It is not the smartest model at the frontier — but it is the smartest free model, embedded in four apps with billions of users, and it signals that Meta is no longer content to be a Llama-only AI company.
For developers: wait for the API and/or open weights. For end users: if you already use WhatsApp or Instagram, Muse Spark is now a credible daily driver, especially for image and multimodal tasks.