April 9th news todayMeta Official launch of the first flagship launched by the SuperIntelligence Labs AI Models Muse Spark.
(a) Numerical support for the joint understanding of images, audio, video and text, and the ability of built-in tools to call, visualize the thinking chain and coordinate with multi-intelligence
(a) Score 42.8 in the HealthBech Hard Open Health Questions and Answers, well above 40.1 in GPT-54, 20.6 in Gemini 3.1 Pro and 14.8 in Opus 4.6; CharXiv Reassoning graph understands score 86.4 and is also leading the competition
At the same performance level, Muse Spark saved 10.3 times more than Llama 4 Maverick Base and 8.2 times more than DeepSeek-V3.1 Base。
But it's also obvious. ARC AGI 2 received only 42.5 points in abstract reasoning puzzles, which lags far behind Gemini 3.1 Pro ' s 76.5 and GPT-5.4 ' s 76.1; Terminal-Bench 2.0 end-coded tasks scored 59.0 and behind GPT-5 ' s 75.1 and Gemini ' s 3.1 Pro ' s 68.5。
There was also an episode in the publication process. Meta, in its assessment graphs, was able to highlight the performance of the home model in an attempt to create a full-fledged, leading vision, which was then criticized by the outside world and directly qualified as a `chart crime' by some netizens. Chief AI Alexander Wang then made a public apology。
Meta indicates that the next generation model (internal designator "Watermelon") is already under development and coding capacity will be the focus of the improvement direction。