Meta launches Muse Wonder: new AI models compared

Meta has presented its new AI models, Muse Wonder and Muse Wonder Full. These models were developed by Meta Superintelligence Labs, led by Alexandr Wang, founder of Scale AI. In recent independent tests by Artificial Analysis, it appears that Muse Wonder Full can compete with the best models from rivals like Anthropic and Google in programming and terminal tasks, but it performs less well in terms of factual knowledge.

Independent benchmarks are essential for evaluating AI models. Muse Wonder Full scored 56 points on the Intelligence Index, while Muse Wonder scored 52 points, compared to the 48 points of its predecessor, Muse Spark 1.3. In relevant categories such as programming and terminal work, Muse Wonder Full achieves scores equal to Claude Code from Anthropic, but lags significantly in factual knowledge. This highlights the weakness of Meta's new models across the broader spectrum of AI skills.

With innovative features such as 'Active Memory', which allows the model to retain information over longer periods, Meta is taking a step towards more advanced AI assistants. The models score around 85.5 on the self-developed MemoryBench, placing them competitively against other players, although these figures come from Meta itself. However, a challenge for Meta remains the hallucination rate, which is 23 percent in these models.

Muse Wonder is accessible via Meta's API for developers, while Muse Wonder Full is only available to Muse Code subscribers. Pricing is comparable to that of competitors, although Meta has not yet indicated when the models will be available in the European Union. With the introduction of Muse Wonder, Meta is back in the race with other major AI players and can further develop towards personal superintelligence in the future.

Read the full article from AI Insider.