Meta’s Muse Spark model is the company’s first true “frontier‑class” AI system built under its new Superintelligence Labs, a natively multimodal reasoning model designed to power Meta AI across Facebook, Instagram, WhatsApp and beyond, with a focus on personal, tool‑using, agentic assistants rather than just static chatbots. It marks a strategic shift: after years of open‑weight Llama releases, Meta has opted to ship a closed, high‑end model that it says nearly matches top systems from OpenAI, Google and Anthropic in language and vision, while still playing catch‑up in coding and complex agentic work.
What Muse Spark is and why Meta built it
Muse Spark is Meta’s new flagship large model, introduced in April 2026 as the first release from Meta Superintelligence Labs (MSL), a high‑profile team formed after CEO Mark Zuckerberg grew dissatisfied with how Llama‑based systems were lagging behind ChatGPT and Claude.
Meta’s initial blog describes Muse Spark as “a natively multimodal reasoning model with support for tool‑use, visual chain of thought, and multi‑agent orchestration,” and says it is built to scale toward “personal superintelligence”, AI that knows your context, can act on your behalf across devices and services, and can reason over long‑form content.
Unlike Llama, Muse Spark is a closed frontier model: Meta has not published its weights or exact parameter count, and early access was limited to the Meta AI app, meta.ai website and selected partners. The company says the first version is designed to be “compact and fast” while still handling complex questions in science, mathematics, and health, and to serve as a base for successive releases.
Capabilities: multimodal, tool‑using, and agentic
From the outset, Muse Spark is built for multimodal reasoning and agentic behavior.
Meta’s technical write‑ups and external coverage highlight several key capabilities:
- Text + vision: Muse Spark accepts both text and image input and produces text output, with strong performance on visual STEM questions and tasks like counting objects, identifying components, and estimating things like meal calories from photos.
- Tool use and computer control: Muse Spark can call tools, use web browsers, and operate virtual computers, enabling it to, for example, click through a user interface, run commands, or help troubleshoot home appliances using on‑screen prompts.
- Visual chain of thought: The model can display intermediate visual reasoning steps, overlaying images (like placing a mug on a shelf in augmented reality) or annotating diagrams as it works through a problem.
- Multi‑agent orchestration (Contemplating mode): Meta is rolling out a Contemplating mode in which Muse Spark spins up multiple agents in parallel to tackle harder tasks. One agent might design a family vacation itinerary while another searches for kid‑friendly activities, with outputs combined into a single plan.
Muse Spark 1.1, announced in July, further sharpens these features, with Meta AI and DataCamp describing the update as “built for agentic tasks” and delivering major gains in tool and computer use, coding, and multimodal understanding.
Benchmarks: where Muse Spark stands in the frontier race
Independent benchmarking suggests Muse Spark is competitive but not top‑of‑the‑table.
Artificial Analysis, which evaluated the model with early access, reports that Muse Spark scores 52 on its Intelligence Index, placing it fourth overall, behind Google’s Gemini 3.1 Pro, OpenAI’s GPT‑5.4 and Anthropic’s Claude Opus 4.6, but ahead of Claude Sonnet 4.6, GLM‑5.1, Grok 4.2 and others. That score marks a huge jump from Meta’s last frontier‑adjacent release: Llama 4 Maverick and Scout scored 18 and 13 on the same index as non‑reasoning models.
On specific benchmarks, Artificial Analysis and Meta highlight:
- Vision: Muse Spark is the second‑most capable vision model they’ve tested, scoring 80.5% on MMMU‑Pro, behind only Gemini 3.1 Pro (82.4%).
- Reasoning and instruction following: It scores 39.9% on HLE (hard language evaluations), trailing Gemini 3.1 Pro and GPT‑5.4 but beating most other models. It also ranks fifth on CritPT, a physics research‑question eval, with 11%, well above Claude Sonnet 4.6 at 3%.
- Token efficiency: Muse Spark used 58 million output tokens for the Intelligence Index, close to Gemini 3.1 Pro (57M) and far lower than GPT‑5.4 (120M) or Claude Opus 4.6 under high‑effort settings.
- Agentic tasks: On GDPval‑AA (real‑world work tasks), Muse Spark scores 1427, behind Claude Sonnet 4.6 (1648) and GPT‑5.4 (1676) but ahead of Gemini 3.1 Pro (1320). On TerminalBench Hard, it trails leading models on command‑line and tool‑use tasks.
The Guardian’s reporting echoes this picture: Muse Spark “was competitive with models from OpenAI, Google and Anthropic in language, but lagged in coding.” Ars Technica notes that agentic performance “does not stand out” yet, though Meta expects Contemplating mode and further reinforcement‑learning (RL) steps to narrow the gap.
Safety, scaling and “thought compression”
Meta is using Muse Spark to showcase its updated thinking on scaling and safety.
Ars Technica reports that the company has tied the launch to an update of its Advanced Scaling Framework, which now explicitly covers a wider array of model risks. Meta says Muse Spark remains “within safe limits across all frontier risk categories assessed,” and promises more detail in a forthcoming Safety & Preparedness Report.
On the training side, Meta emphasizes that Muse Spark shows “consistent and predictable improvements” from extra RL steps after pretraining, addressing criticism that earlier Llama models didn’t leverage RL effectively. The RL regime includes “thinking time penalties”, costs applied to long reasoning traces, intended to balance accuracy with token efficiency. That’s part of what external analysts have dubbed “thought compression”: spending more test‑time on reasoning without exploding latency or cost.
The multi‑agent Contemplating mode is central to this approach. Meta says it can coordinate up to 16 agents reasoning simultaneously, allowing the system to tackle harder problems while maintaining similar user‑perceived latency. For developers and users, this is where Muse Spark is supposed to evolve from a chatbot into a true “agentic” system.
How and where you can use Muse Spark
Right now, Muse Spark is mainly accessible inside Meta’s own ecosystem.
The initial release put Muse Spark behind the Meta AI assistant on the meta.ai website and in the Meta AI app, where users can chat, interrupt, switch topics or swap languages as they talk. Meta says the model will replace existing Llama‑based assistants across WhatsApp, Instagram, Facebook, and its Ray‑Ban Meta smart glasses in the weeks after launch.
Users must log in with a Meta account (Facebook, Instagram, etc.) to access Muse Spark, and while Meta doesn’t explicitly say their account data is used for training, it generally trains on public user content and has framed Muse Spark as a “personal” model that can adapt to users over time.
On the developer side, Muse Spark initially shipped without a public API; Meta offered only a private preview of a Meta Model API to select partners. In July, the company announced Muse Spark 1.1 and a Meta Model API designed for agentic workloads, with a 1‑million‑token context window and stronger coding and computer‑use capabilities. DataCamp calls the new API “Meta’s agentic model and API,” aimed at developers building tools, coding assistants and multimodal apps.
How Muse Spark fits into Meta’s broader AI strategy
Muse Spark is as much about Meta’s business and platform strategy as raw model performance.
Fortune reports that the release marks Meta’s “first new AI model since Llama 4,” and that the company is using it as part of a “ground‑up overhaul” of its AI efforts. Meta has historically open‑sourced Llama models, helping them spread across the industry, but with Muse Spark it has chosen a closed approach, limiting access to its own products and selected partners.
The New York Times notes that Muse Spark represents Zuckerberg’s attempt to “catch up to AI leaders Anthropic, Google and OpenAI” after roughly a year of relative quiet. By integrating the model deeply into Facebook, Instagram, Threads, WhatsApp and wearables, Meta is betting that control over distribution, billions of users, and their daily habits, can compensate for trailing slightly on some benchmarks.
At the same time, Muse Spark signals a more product‑centric design than Llama. Rather than being marketed as a general‑purpose model for anyone to download, it is framed as a “personal superintelligence” tuned for conversational, visual, health and everyday planning tasks inside Meta’s platforms.
What to watch next
Understanding Muse Spark today is partly about watching where it goes next.
Meta has already teased future iterations with larger context windows (1M in Muse Spark 1.1), stronger coding and agentic performance, and wider API access. It has also suggested that some components or smaller variants could eventually be open‑sourced, though no timelines are firm.
For users, the biggest changes may be subtle: smarter Meta AI in feeds and messages, more capable visual assistance in apps and on smart glasses, and new multi‑agent features that quietly structure trips, projects, or routines. For developers and competitors, Muse Spark is a signal that Meta is no longer treating open‑source LLMs as its only, it wants a proprietary frontier model, too.
In that sense, understanding Muse Spark is understanding Meta’s bid to be both a platform and a frontier AI player at the same time.