Meta’s Muse Spark tested

Meta just put a new model called Muse Spark inside the Meta AI app and on meta.ai, and the early tests people are sharing are less about chat and more about doing work: writing code, reading images, and handling longer reasoning in one place. A multimodal model is an artificial intelligence system that can take in more than plain text, the way a person can read a note and also look at a photo. Meta says Muse Spark accepts text and vision input, so a user can type a prompt or upload an image instead of describing everything by hand. Reasoning is the part where the model does more than autocomplete a sentence and instead works through a problem step by step, like showing scratch paper before giving an answer. Meta says Muse Spark was built to handle harder questions in science, math, and health, and TechCrunch reports that Meta is also adding a “Contemplating” mode for more complex tasks. Agents are small helper processes that split up a job, like sending three interns to check three parts of the same story at once. Meta says Muse Spark can launch multiple subagents in parallel, which is how it tries to spend more test-time reasoning without making users wait as long. That mix matters because coding, image reading, and reasoning usually live in separate tools. In the April 9 YouTube test, the reviewer ran Muse Spark through coding tasks, multimodal prompts, and agent workflows in a single session and described it as an all-rounder rather than a narrow demo. Meta launched Muse Spark on April 8, 2026, and called it the first model from Meta Superintelligence Labs, the reorganized group now leading its frontier artificial intelligence work. TechCrunch says that lab was formed after Meta fell behind OpenAI and Anthropic in public perception, and that former Scale AI chief executive Alexandr Wang was recruited to lead it. The benchmark picture says Muse Spark is competitive, but not the top model overall. Artificial Analysis says it scored 52 on its Intelligence Index, behind Gemini 3.1 Pro and GPT-5.4 at 57 and behind Claude Opus 4.6 at 53, while still landing in the top five models it has tested. The stronger surprise is vision. Artificial Analysis says Muse Spark scored 80.5 percent on the MMMU-Pro vision benchmark, which it describes as the second-best vision result it has measured, behind only Gemini 3.1 Pro Preview at 82.4 percent. The weaker spot is agentic work on real jobs, which is the category closest to “can this actually help me finish something.” Artificial Analysis says Muse Spark scored 1427 on its GDPval-AA work-task evaluation, behind Claude Sonnet 4.6 at 1648 and GPT-5.4 at 1676, so the product looks stronger at understanding than at fully autonomous execution. Meta is also changing how it ships models. Artificial Analysis says Muse Spark is Meta’s first frontier model that is not released as open weights, and Meta says broader access will come first through its own products and then through a private-preview application programming interface for selected partners. For a newsroom, that means one interface can draft a script, inspect a screenshot, pull structured details from an image, and help write the automation around that workflow. Muse Spark is not the best model on every leaderboard, but the April 2026 tests suggest Meta’s bigger shift is bundling several useful abilities into one fast system that is already live inside consumer products.