BridgeMind On Muse Wonder: I've Never Seen A Model This Benchmaxxed In My Life
Summary
A critique of Meta's Muse Wonder model, arguing it is heavily 'benchmaxxed' — tuned to top key benchmarks but performing poorly on real-world tests like the BridgeBench lava lamp task, and falling short of frontier capabilities.
Similar Articles
Muse Wonder 1.0 Benchmarks
Meta's Muse Wonder 1.0, described as a Fable Class Model, is benchmarked in a shared comparison chart.
Meta releases its first frontier model, Muse Wonder
Meta's Superintelligence Lab released its first frontier models, Muse Wonder and Muse Wonder Full, claiming benchmark performance comparable to Fable 5.1 and GPT-6 Astra, with a special emphasis on long-term 'active memory' and strong MemoryBench results.
@MaximeRivest: muse wonder is beating all other models at pointing things in images. what a come back from meta !!
A tweet praising Meta's Muse Wonder model for outperforming other models at pointing to objects in images, marking a notable comeback for Meta in visual grounding.
Muse Wonder
Meta 今日发布 Muse Wonder 系列前沿模型(Muse Wonder 与 Muse Wonder Full),主打个人与专业智能体任务,具备 Active Memory 与 Active Focus 能力,已在编程、法律、金融等领域基准测试中达到领先水平,将通过 Meta Model API 和 Muse Code 提供服务。
Meta Muse Wonder: pricing, benchmarks and limits - MACCUS
Meta released Muse Wonder and Muse Wonder Full, its most capable AI models yet, designed for personal and professional agents, available via the Meta Model API and Muse Code subscriptions. The article details pricing ($4/$20 per million tokens), benchmarks versus Claude Opus 5.5, and limits like Contributor mode and restricted availability of the Full variant.