{"id":382057,"date":"2026-04-08T21:27:07","date_gmt":"2026-04-08T21:27:07","guid":{"rendered":"https:\/\/www.newsbeep.com\/il\/382057\/"},"modified":"2026-04-08T21:27:07","modified_gmt":"2026-04-08T21:27:07","slug":"metas-superintelligence-lab-unveils-its-first-public-model-muse-spark","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/il\/382057\/","title":{"rendered":"Meta&#8217;s Superintelligence Lab unveils its first public model, Muse Spark"},"content":{"rendered":"<p>In that same post, Meta also touts \u201cContemplating\u201d mode, which it says will be \u201crolling out gradually\u201d and which can \u201corchestrate multiple agents that reason in parallel.\u201d By using up to 16 agents thinking in concert simultaneously, Meta says that Contemplating mode \u201cenables superior performance with comparable latency.\u201d That \u201csuperior performance\u201d includes a reported high water mark of 58.4 on <a href=\"https:\/\/agi.safe.ai\/\" rel=\"nofollow noopener\" target=\"_blank\">Humanity\u2019s Last Exam<\/a> (with the use of external tools), according to Meta.<\/p>\n<p>\n      A Meta graph shows how additional training leads to \u201ccompression\u201d of token usage before additional gains in accuracy.\n    <\/p>\n<p>\n      A Meta graph shows how additional training leads to \u201ccompression\u201d of token usage before additional gains in accuracy.<\/p>\n<p>And while previous Llama models <a href=\"https:\/\/x.com\/burkov\/status\/1908961952141091196\" rel=\"nofollow\">faced criticism for not taking advantage of reinforcement learning<\/a>, Meta says Muse Spark shows \u201csmooth predictable gains\u201d after additional RL steps after pretraining, \u201cimproving model reliability without compromising reasoning diversity.\u201d That reinforcement learning system also makes use of \u201cthinking time penalties,\u201d which Meta says balance the need to \u201cmaximize correctness\u201d with optimizing the number of tokens used. In testing on <a href=\"https:\/\/www.vals.ai\/benchmarks\/aime\" rel=\"nofollow noopener\" target=\"_blank\">the AIME 2025 benchmark<\/a>, Meta says it saw a \u201cphase transition\u201d where the model started compressing equally accurate reasoning into \u201csignificantly fewer tokens.\u201d After that compression, subsequent trained models slowly increased the token usage again to achieve even higher accuracy in less overall time than the previous uncompressed versions.<\/p>\n<p>The release of Muse Spark comes alongside <a href=\"https:\/\/ai.meta.com\/blog\/scaling-how-we-build-test-advanced-ai\/\" rel=\"nofollow noopener\" target=\"_blank\">an update to Meta\u2019s Advanced AI Scaling Framework<\/a>, which the company says now covers a broader range of potential AI risks. The company says that the model \u201cfalls within safe margins across all frontier risk categories we measured,\u201d but says that more details will only be available in an upcoming Safety &amp; Preparedness Report.<\/p>\n<p>Muse Spark is available now in the Meta AI app and <a href=\"https:\/\/meta.ai\/\" rel=\"nofollow noopener\" target=\"_blank\">via the meta.ai<\/a> website, as well as a private preview API for \u201cselect partners.\u201d Meta says the model will be available via WhatsApp, Instagram, Facebook, Messenger, and AI glasses \u201cin the coming weeks.\u201d<\/p>\n","protected":false},"excerpt":{"rendered":"In that same post, Meta also touts \u201cContemplating\u201d mode, which it says will be \u201crolling out gradually\u201d and&hellip;\n","protected":false},"author":2,"featured_media":382058,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[345,343,344,85,46,125],"class_list":{"0":"post-382057","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-artificial-intelligence","8":"tag-ai","9":"tag-artificial-intelligence","10":"tag-artificialintelligence","11":"tag-il","12":"tag-israel","13":"tag-technology"},"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/posts\/382057","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/comments?post=382057"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/posts\/382057\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/media\/382058"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/media?parent=382057"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/categories?post=382057"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/il\/wp-json\/wp\/v2\/tags?post=382057"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}