Meta：此次Muse Spark能否点燃投资者对其人工智能叙事的信心？

Mark Shmulik+1 917 344 8508mark.shmulik@bernsteinsg.com U.S. Internet Meta Platforms Inc. Wenhuan Chang+1 917 344 8546wenhuan.chang@bernsteinsg.com RatingOutperform Deeksha Pandey+1 917 344 8447deeksha.pandey@bernsteinsg.com Price Target 900.00 USD META Meta: Will this Muse Spark investor belief in their AI story? Sandwiched between Anthropic’s non-launch of Mythos and what we expect to be OpenAI’sresponse shortly, Meta found a window to launch their own frontier model, Muse Spark, thatcomfortably surpassed lukewarm investor expectations on model performance.Our earlymodel tests were quite impressive, but as we’ve noted a high-performing model is table-stakes, now comes the hard and/or fun part of shipping products and features that win overconsumers, advertisers, and businesses against the plethora of compelling alternatives in themarket. Meta’s invited to the Kumite, but can Jean Claude Van Zuck win the tournament? Investment Implications We rate Meta Outperform. DETAILS EXHIBIT 3:Opportunity & use cases across 3B+ users, 200M+ creators, 200M+ businesses, 10M+ advertisers, and~79K employees META’S MUSE SPARK MODEL HAS ARRIVED Meta announced today the availability of its new family of models (“Muse”) and the first AI model release of this family (“MuseSpark”) from Meta’s Superintelligence Labs division formed nine months ago. Muse Spark is the first frontier-class model fromMeta since Llama 4 Maverick was released in April 2025. •Meta is releasing its first model, Spark (as part of a broader suite of models called Muse) that will power Meta AI, thecompany’s internally developed AI assistant. Spark is built from the ground up to integrate visual information acrossdomains and tools, and is natively multimodal (can process images, video and audio) with support for tool-use, visual chain ofthought, and multi-agent orchestration.It currently only powers Meta AI within the Meta AI app, but will be rolling out toWhatsApp, Instagram, Facebook, Messenger, and AI glasses in the coming weeks. •Meta gets a spot in the Top 5 with better than expected model performance… for nowMeta has closed the gap withfrontier models with Muse Spark, now in range of Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 on the Artificial AnalysisIntelligence Index (see Exhibit 4; the Artificial Analysis team was given early access by Meta to independently benchmark themodel). Notably, Spark lands ahead of Claude Sonnet 4.6, GLM-5.1, MiniMax-M2.7, Grok 4.20 in their assessment. We notethat Anthropic’s Mythos and OpenAI’s next models should once again raise the frontier bar when released for GA. •Notably token efficient.Muse Spark is notably token efficient relative to its intelligence level (see Exhibit 6 and Exhibit 7). Itrequired 58M output tokens to run the Artificial Analysis Intelligence Index, broadly in line with Gemini 3.1 Pro Preview (57M),while consuming materially fewer tokens than other frontier models such as Claude Opus 4.6 (Adaptive Reasoning, maxeffort, 157M), GPT-5.4 (xhigh, 120M), and GLM-5 (110M). Meta highlights that a rebuilt pretraining stack spanning modelarchitecture, optimization, and data curation has enabled Muse Spark to reach comparable capabilities with “over an order ofmagnitude less compute” than its prior model, Llama 4 Maverick. •Scores particularly high on Multimodal, Vision.The Artificial Analysis team noted that Muse Spark was the second-most capable vision model they have benchmarked (it scores 80.5% on MMMU-Pro, behind only Gemini 3.1 Pro Preview(82.4%)). In his Threads post, CEO Mark Zuckerberg noted that Spark is “particularly strong in areas related to personalsuperintelligence” including visual understanding, health, social content, shopping, games, and more. Comparison acrossbenchmarks shows Spark’s impressive performance in multimodal perception, reasoning, health, and agentic tasks (seeExhibit 8). Contemplating mode (that orchestrates multiple agents that reason in parallel, enabling deeper, more exhaustiveproblem-solving) allows a material uplift in performance on difficult benchmarks vs key frontier models (see Exhibit 9).Contemplating mode will be rolling out gradually in meta.ai. Long-horizon agentic systems and coding workflows are modelimprovement areas. •Especially impressive on shopping research; reasonably competitive on agentic.We tested Meta AI’s shoppingfeatures first-hand, and were very impressed. Meta highlighted that Muse Spark is purpose-built for Meta’s products andover time, will unlock features that cite recommendations and content people share across Instagram, Facebook, andThreads. Alex Wang (Chief AI officer at Meta) posted on Threads about shopping mode making recommendations basedon creators, brands, and styling content across apps as well as Meta AI pulling real conversations across Meta apps (seeExhibit 11). Muse Spark also seems to be reasonably competitive on agentic workflows, which matters given its likely role inpowering Manus AI. •Trained on Blac