Muse Spark is the latest model from Meta. Everyone is talking about benchmarks and Mark Zuckerberg’s billion-dollar bet, but in this post, I want to focus on what this means for developers building AI systems. Image by author Muse Spark is a general-purpose model, and its most interesting feature is Contemplating mode, an inference-time multi-agent system that runs multiple agents in parallel to solve complex problems without proportionally increasing latency. This is an architectural shift from the existing models, where each agent solves a problem sequentially and builds upon its previous outputs, to multiple agents working simultaneously, and the agents’ outputs are refined and aggregated to return the answer to the user. This shift has real implications for anyone who is building a multi-agent system. This post covers Contemplating mode, and what is happening under the hood: the tradeoffs of parallel and sequential execution, what Meta got right, and where the gaps remain in agentic solution development. Background Muse Spark was developed out of the immense pressure Meta faces to establish itself as a serious AI contender. About a year ago, Meta released its Llama 4 model, which received an underwhelming response from the developer community, with many reporting that its benchmark results had been manipulated using specialized sub-models never released to the public. To add to Meta’s problems, Anthropic and OpenAI continued releasing new and better models. Mark Zuckerberg responded to growing tension by hiring Alexandr Wang, the co-founder and former CEO of Scale AI, as Chief AI Officer, backed by a $14.3 billion investment in Scale AI for a 49% stake. Zuckerberg assembled Meta Superintelligence Labs over the next nine months, staffed with researchers poached from Google, Anthropic, and OpenAI. The team rebuilt everything from scratch, improving model architecture, optimization pipelines, and data curation. Muse Spark is the first model to emerge from this effort. Unlike Llama 4, this model is proprietary rather than open-source. This marks a significant strategy shift for a company that built its reputation on open-source releases like Llama. Meta’s AI capital expenditures in 2026 are projected at $115–135 billion, nearly double last year’s level. The stakes could not be higher. What Is Muse Spark? Muse Spark is a general-purpose model like ChatGPT and Claude, but unlike these models, which were originally built for text, with vision added later, Muse Spark was built to be multimodal from day one. It integrates text, voice, and images into one system from the start rather than as an add-on capability. Muse Spark is designed to be small, fast, and scalable. It is capable of reasoning through complex problems in math, health, and science. The model is currently available on meta.ai and the Meta AI app and is rolling out to Instagram, Facebook, WhatsApp, and Ray-Ban Meta AI glasses. Contemplating Mode: Under The Hood Traditionally, language models solve problems sequentially, each step building on the previous one. This sequential processing increases latency: the more complex the problems, the longer the wait time. Contemplating mode takes a different approach: think of it like hiring a team to come up with a solution to a given problem. Each team member works independently on sub-tasks and contributes to the outcome. The main advantage of this architecture is that you tackle complex problems without increasing the latency. There are three stages in this architecture: Solution generation: Multiple agents independently attack the problem. Iterative self-refinement: Each agent iteratively refines the solution that they can generate. Aggregation: Outputs are combined for a final, higher-quality answer. Benchmark Results: Humanity’s Last Exam With Tools, a benchmark testing the expert-level multidisciplinary knowledge: Muse Spark scored 50.2, ahead of Gemini 3.1's score of 43.9. FrontierScience Research, Muse Spark scored 38.3, ahead of GPT-5.4 Pro, which scored 36.7. Image from Meta Official Blog This shows Muse Spark is not just competitive as a general-purpose model, but also demonstrates domain expertise in specific areas. It is worth noting that Contemplating mode is an inference-time feature, and the underlying model does not change. Meta is scaling at the moment when the answer is being generated, not during the training phase. What Does It Mean for a Developer? As someone who builds multi-agent systems professionally, here is my take on Contemplating mode: Parallel processing is not always better : Parallel processing is possible only when a problem can be broken down into multiple independent tasks, for example, summarizing multiple documents, but if you are generating a lab report where each step depends on the previous one and the result of the first step affects the output of the next step, then sequential execution is the better choice. Aggregation of the outputs: Meta did not talk much about how the aggregation is happening, whether by taking a majority vote, or scoring and selecting the best output. This is an important step, as this is the final answer that is returned to the user. Get this wrong, and low latency would be meaningless. Thought Compression: Meta trained the model with a penalty on the thinking time , so it is trained not just on correctness but also using fewer tokens to solve a problem. This reduces inference cost and latency. This highlights something important: the verbose chain-of-thought reasoning is expensive, and rewarding brevity without sacrificing accuracy is a meaningful training insight. Image from Meta Official Blog Where Does Muse Spark Fall Short? The model is no longer open-source, unlike LLama, whose weights were publicly released. The user needs to log in to a Meta account to use the model, which raises serious concerns about data privacy, especially when used in an enterprise setting handling sensitive data. On coding, Muse Spark is competitive, but not leading on the SWE benchmark; it scores 77.4, behind Claude Opus 4.6 Max at 80.8 and Gemini 3.1 Pro High at 80.6. On abstract reasoning, Muse Spark scored 42.5 on ARC-AGI2, significantly behind Gemini 3.1 Pro at 76.5. This suggests that the model is not good at solving novel problems and struggles when generalization is needed, rather than pattern recognition from data. Final Thoughts Zuckerberg wrote on Threads that Meta’s goal is to build AI products that “don’t just answer your questions but act as agents that do things for you.” Muse Spark is the first concrete step in that direction.[Threads, April 2026] Whether Meta can close the gap with OpenAI and Anthropic is an open question. The benchmark results are uneven, and by moving away from open-source, Meta has introduced a trust deficit with the developer community that will take time to rebuild. But Contemplating mode points to something genuinely important regardless of who built it. The biggest takeaway is that it is not about larger models or longer training runs. It is about having a smarter generation of answers, and that shifts the implications of how you design an agent today. If you are building an agent using Muse Spark, think about how the parallel execution can help you, and audit every aggregation step. The gap between a well-orchestrated agent and a poorly designed agent is growing, and Contemplating mode is an example of where the frontier is headed. Muse Spark may not be the best model out there, but the idea is definitely new and something to think on. A message from our Founder Hey, Sunil here. I wanted to take a moment to thank you for reading until the end and for being a part of this community. Did you know that our team run these publications as a volunteer effort to over 3.5m monthly readers? We don’t receive any funding, we do this to support the community. If you want to show some love, please take a moment to follow me on LinkedIn , TikTok , Instagram . You can also subscribe to our weekly newsletter . And before you go, don’t forget to clap and follow the writer️! A Deep Dive into Muse Spark was originally published in Artificial Intelligence in Plain English on Medium, where people are continuing the conversation by highlighting and responding to this story.