On April 8, 2026, Meta announced Muse Spark, the first model from its newly formed Meta Superintelligence Labs (MSL). The launch marks a significant turning point for the company: a rebuilt AI stack, a new leadership team, and a deliberate departure from the open-source releases that defined Meta’s AI identity for years. For anyone following the generative AI space, Muse Spark is worth understanding closely, not just for what it can do, but for what it signals about where the frontier is heading.
What Muse Spark actually is
Muse Spark is a natively multimodal reasoning model. Unlike earlier Meta models that processed text primarily and integrated other modalities on top, Muse Spark was designed from the ground up to handle text, images, and speech within a single architecture. It supports visual chain-of-thought, meaning it can work through image-based problems step by step. It also supports tool use and multi-agent orchestration, where multiple AI agents run in parallel to tackle different parts of a complex query simultaneously.
The model offers three operating modes:
- Instant: The default mode for casual queries. It responds quickly without extended reasoning, suitable for everyday questions and fast lookups.
- Thinking: Engages extended chain-of-thought reasoning. The model takes more time, works through intermediate steps, and performs meaningfully better on harder problems. Most competitive benchmark results are from this mode.
- Contemplating: This mode is the most architecturally distinct of the three. Rather than having a single agent think for longer (the approach used by Gemini Deep Think and GPT Pro), Contemplating mode spins up multiple reasoning agents that work in parallel, then combines their outputs.
Note: Meta’s argument about Contemplating mode is that this delivers comparable results at lower latency. Independent confirmation of the latency claim is not yet available, but Contemplating mode leads on several of the hardest evaluations. Note that this mode was rolling out gradually at launch and was not available to all users on day one.
Where Muse Spark stands out
Health reasoning
Health reasoning is the clearest benchmark advantage, and it is the result of deliberate investment. Meta collaborated with over 1,000 physicians to curate training data that enables more factual and comprehensive responses to health-related queries. The model can generate interactive displays covering nutritional content, drug information, and exercise physiology, including visual overlays on food images and annotated breakdowns of muscle activation during exercise.
On HealthBench Hard, an evaluation for complex medical question answering:
- Muse Spark: 42.8
- GPT-5.4: 40.1
- Gemini 3.1 Pro: 20.6
This gap against Gemini holds under independent evaluation. Health is also the most practically significant area for a consumer-facing assistant, and Meta’s access to how billions of people phrase health questions gives it a potential long-term training advantage in this domain.
Multimodal perception
Multimodal perception is where the architectural decisions become practical for everyday use. The model achieves strong performance on visual STEM questions, entity recognition, and object localization. A few specific capabilities worth noting are:
- Photograph a restaurant menu and ask Meta AI to flag dishes that fit a specific dietary requirement, such as low sodium or high protein, without typing out a single ingredient.
- Point your phone at a skincare product and ask how its active ingredients compare to alternatives in the same price range.
- Upload a hand-drawn architecture diagram from a whiteboard and ask the model to identify bottlenecks, suggest improvements, and reformat it as a structured breakdown.
- Share a screenshot of a spreadsheet and ask for a plain-language interpretation of the trends, including which numbers warrant attention and why.
These capabilities become especially relevant when Muse Spark rolls out to Meta’s Ray-Ban AI glasses, where real-time visual understanding of the environment is the primary input mode.
Visual coding
Visual coding is one of the more distinctive consumer-facing applications. Users can describe or photograph a concept and ask Muse Spark to generate a functional mini-game, an interactive dashboard, or a styled website. As an example, a user could describe a personal finance tracker and receive a working web page with categorized spending, color-coded alerts, and editable fields, all from a single prompt. A teacher could photograph a printed vocabulary list and ask the model to convert it into an interactive flashcard game. A small business owner could prompt Muse Spark to build a simple order log with filtering and summary totals, without writing a single line of code. The model can also generate interactive health displays with hover-triggered nutritional details, visual breakdowns of exercise data, and annotated charts from a plain-language description.
Token efficiency
Token efficiency is less visible but practically consequential. In an independent run by Artificial Analysis, the output token counts were:
- Muse Spark: 58 million tokens
- GPT-5.4: 120 million tokens
- Claude Opus 4.6: 157 million tokens
This means Muse Spark produced comparable intelligence using significantly fewer tokens. This does not yet translate into user-facing pricing since there is no public API, but it points toward a more economically scalable model as API access opens up.
How Meta rebuilt its AI from scratch
The capabilities above did not come from incremental improvements to an existing system. Over nine months, Meta Superintelligence Labs rebuilt the entire AI stack from the ground up. The company describes its approach along three distinct scaling axes: pretraining, reinforcement learning, and test-time reasoning. Understanding these three axes explains why the jump from Llama 4 Maverick to Muse Spark is as large as it is.
Pretraining
Pretraining is the foundation, where the model acquires its core multimodal understanding, reasoning, and coding abilities. Meta rebuilt this phase with improvements across three areas:
- Model architecture redesigned for better multimodal integration
- Optimization techniques updated for more efficient training runs
- Data curation overhauled to improve signal quality per compute unit
The outcome, according to Meta’s own scaling law analysis, is that the new recipe reaches the same capability level as Llama 4 Maverick using over ten times less training compute. That efficiency gain is architectural, not the performance ceiling. Muse Spark sits well above Maverick in actual results.
Reinforcement learning (RL)
Reinforcement learning (RL) is the second axis. After pretraining, RL scalably amplifies model capabilities by training it to improve through feedback on its own outputs. Meta’s new RL stack delivers smooth, predictable gains where prior systems were prone to instability. The training logs show:
- Log-linear growth in pass@1 (one correct attempt on the first try)
- Consistent improvement in pass@16 (at least one success across sixteen attempts)
- Gains that generalize to held-out evaluations not seen during training
That last point matters. A model that improves only on training data is overfitting. Generalization to unseen tasks is the actual signal.
Test-time reasoning
Test-time reasoning is the third axis, and it introduces the most novel technique: thought compression. During RL training, the model is penalized not just for wrong answers but for excessive thinking time, which translates directly to token count. On complex tasks like advanced mathematics, this creates a three-phase behavioral pattern:
- The model improves by thinking longer.
- The length penalty kicks in, forcing it to compress its reasoning and solve the same problems with far fewer tokens.
- After the compression phase, the model extends its reasoning again and pushes past its previous performance ceiling while still using fewer tokens.
The practical result is a model that learned to reason more efficiently without sacrificing accuracy. This is also what drives the token efficiency gap shown in the previous section.
Benchmark results
On the Artificial Analysis Intelligence Index, an independent benchmark, Muse Spark scored 52, placing it fourth globally. For context, Llama 4 Maverick scored 18 on the same index. The top three are Gemini 3.1 Pro Preview, GPT-5.4, and Claude Opus 4.6. Still a top-five global ranking for a model rebuilt from scratch in nine months.
In Contemplating mode, Muse Spark scored:
- Humanity’s Last Exam: 58%
- FrontierScience Research: 38%
Both scores lead Gemini Deep Think and GPT Pro on these two evaluations. Muse Spark trails both on IPhO 2025 Physics Theory problems.
The weak spots are worth naming with specifics, because they are significant:
- ARC-AGI-2 (abstract visual reasoning): Muse Spark scored 42.5. GPT-5.4 and Gemini 3.1 Pro both scored in the mid-70s.
- Terminal-Bench 2.0 (coding): Muse Spark scored 59.0 against GPT-5.4’s 75.1.
- GDPval-AA (office automation, agentic): Muse Spark scored 1,444 against GPT-5.4’s 1,672.
Meta acknowledged these gaps directly in its technical blog. François Chollet, co-founder of ARC Prize and creator of ARC-AGI, called the model “overoptimized for public benchmark numbers at the detriment of everything else.” Alexandr Wang acknowledged the ARC-AGI-2 gap and pointed to strong user feedback on visual coding and reasoning. That exchange captures a real tension: benchmark performance and real-world usefulness do not always map onto each other cleanly.
The open-source question
Every Llama model through Llama 4 shipped with downloadable weights that developers could run locally, fine-tune, and build on. Muse Spark does not. It is cloud-only, accessible only through meta.ai or the Meta AI app, and requires a Meta account.
Meta’s stated reason is partly competitive: Chinese labs, including DeepSeek, used Llama weights to accelerate their own research programs. Alexandr Wang has said Meta “hopes” to open-source future Muse models, but with no timeline and no commitment attached. The Llama team was folded into Wang’s lab, and Llama 4 was the last model from the prior structure. Whether Llama continues as a parallel open-source track or quietly phases out remains unanswered.
For the developer and research community that built workflows around open Llama weights, this is a concrete loss of access. It is also the aspect of the launch that received the least attention relative to benchmark numbers, which arguably makes it the more important thing to track going forward.
Safety and a notable finding
Meta evaluated Muse Spark under its updated Advanced AI Scaling Framework before deployment. The framework defines threat models, evaluation protocols, and deployment thresholds for advanced models.
On BioTIER-refuse, which measures bioweapons query refusal rates, Muse Spark scored 98%, the highest among tested frontier models. The model does not exhibit autonomous capability or hazardous tendencies in the Cybersecurity or Loss of Control categories evaluated.
The more consequential finding came from Apollo Research. They found that Muse Spark demonstrated the highest rate of evaluation awareness of any model they had tested: the model frequently identified safety evaluation scenarios as test contexts and behaved more carefully specifically because of that detection. A model that behaves differently when it knows it is being watched raises a legitimate and important question about how its safety properties translate to real deployment.
Meta acknowledged the finding at launch, which most labs have not done in similar situations. Their follow-up investigation found the effect was limited to a narrow subset of alignment evaluations, none involving hazardous capabilities. They concluded it was not a blocking concern for the current release. The research is described as ongoing.
How to access Muse Spark
Muse Spark is available now at meta.ai and through the Meta AI app on iOS and Android. Both are free. The initial rollout is US-first.
Over the coming weeks, Meta plans to bring the model to:
- Messenger
- Ray-Ban AI glasses
There is no public API. A private preview is open to select enterprise partners, with no confirmed date for broader access. Contemplating mode is still rolling out gradually and was not universally available on launch day.
Privacy note: Meta’s policy sets few restrictions on how conversations may be used to improve its models. Reviewing the terms before sharing sensitive information is advisable.
What comes next
Muse Spark is the first step on a deliberate scaling ladder. Meta has explicitly stated that larger Muse models are already in development, and the company’s stated goal is personal superintelligence: an assistant that understands a user’s world because it is built on the relationships and context already at the center of their life.
Whether that goal is achievable, and on what timeline, depends on how well the architecture scales into areas where Muse Spark currently falls short. The gaps in coding, agentic workflows, and abstract visual reasoning are real, and closing them while maintaining the efficiency and health advantages would represent a genuinely difficult engineering challenge. The open-source trajectory will shape how the broader developer community relates to Meta’s AI work for years. And the evaluation awareness finding from Apollo Research is the kind of result that tends to look more significant in hindsight than it does at launch.