Muse Spark

Overview

Freshness note: Model capabilities, limits, and availability can change quickly. This profile is a point-in-time snapshot last verified on April 8, 2026.

Muse Spark is the first model in Meta’s new Muse family and the first public model launch tied directly to Meta Superintelligence Labs. Meta positions it as a natively multimodal reasoning model with tool use, visual chain of thought, and multi-agent orchestration. In product terms, this is not a distant research tease: Meta says Muse Spark is available today at meta.ai and in the Meta AI app, with a private API preview opening to select users.

The important shift is not just that Meta has another frontier model. It is that the launch page frames Muse Spark as the first step on a new scaling ladder and as the first output of a rebuilt stack spanning research, training, and infrastructure. That gives the model more significance than a routine point release. It is Meta’s current signal that its post-Llama frontier direction is productized again.

Capabilities

Meta’s launch claims center on four areas: multimodal perception, reasoning, health, and agentic behavior. The model is meant to understand visual information, work across tools, and support more deliberate reasoning than the ordinary “instant answer” assistant mode people associate with consumer AI chat products.

The most distinctive feature in the launch is Contemplating mode, which Meta describes as multiple agents reasoning in parallel. Meta explicitly positions that mode against extreme-reasoning offerings such as Gemini Deep Think and GPT Pro, and claims 58% on Humanity’s Last Exam and 38% on FrontierScience Research. Even if those numbers need continued independent scrutiny, they show how Meta wants this model understood: not just as a social-app assistant, but as a serious reasoning system with configurable test-time compute.

Meta also emphasizes personal use cases. The launch examples focus on visual STEM reasoning, localizing annotations in the world around you, turning prompts into playable experiences such as a Sudoku game, and supporting health-oriented explanations. Meta says it worked with more than 1,000 physicians on health-related training data to improve factuality and completeness in health reasoning.

Technical Details

Meta describes Muse Spark as a natively multimodal reasoning model, but the current public source set used for this refresh does not expose a clean developer-style spec sheet with a public token window, max output limit, or public API rate card. For that reason, Signal Lens stores contextWindow and maxOutput as 0 and treats them as unavailable in the UI until Meta publishes clearer API documentation.

The more useful technical signal right now is the scaling story Meta chose to publish. The launch post highlights improvements across pretraining, reinforcement learning, and test-time reasoning. Meta says its rebuilt pretraining recipe can reach the same capability level with over an order of magnitude less compute than Llama 4 Maverick, and it also claims smoother reinforcement-learning gains plus thought compression during test-time reasoning. In plain terms, Meta is arguing that Muse Spark is a stack-level reset, not just a bigger checkpoint.

Pricing & Access

Current official availability is:

• meta.ai
• the Meta AI app
• a private API preview for select users

Meta says Contemplating mode will roll out gradually in meta.ai. The current public launch material does not publish public token pricing or a broad self-serve API access path, so teams should treat Muse Spark as a product-first launch with limited developer access rather than as a fully documented commodity API model.

Best Use Cases

Muse Spark is most relevant for multimodal assistant experiences where visual understanding, tool use, and stronger reasoning matter together: visual troubleshooting, interactive learning flows, health explanation interfaces, and consumer-facing assistant experiences that need more than one quick-pass response.

It is less suited to production planning when you need stable API documentation, public pricing, or long-settled enterprise availability. Today, the strongest case for evaluation is “Meta’s newest reasoning direction is now live in its own products,” not “this is already the safest default for third-party deployment.”

Comparisons

• Llama 4 Maverick (Meta): Meta’s own launch framing treats Muse Spark as a major step forward from the Llama 4-era baseline and as the first result of a rebuilt training stack.
• Gemini Deep Think / GPT Pro: Meta explicitly says Contemplating mode is meant to compete with the extreme-reasoning modes of those models on harder tasks, though Meta’s public developer documentation is still thinner than what those rival ecosystems expose.
• Third-party benchmark context: Artificial Analysis reports Muse Spark as near-frontier on overall intelligence and especially strong on vision, while also noting that agentic benchmarks still trail GPT-5.4 and Claude Sonnet 4.6. That should be read as external context, not as the primary basis for availability or product claims.