# Mark Zuckerberg's Dual-Track AI: Llama 5 Open, Muse Spark Closed _Meta simultaneously shipped Llama 5 as open-weights and Muse Spark as a closed model on April 8, 2026, splitting its AI strategy for the first time while capex guidance hit $115-135 billion for the year._ **Published:** 2026-06-05 **Source:** https://www.startuphub.ai/ai-news/ai-figures/2026/figure-mark-zuckerberg-dual-track-open-closed-2026-06-05 --- **Meta simultaneously shipped two AI models on April 8, 2026, a structural first for the company: Llama 5 as open-weights and Muse Spark as a closed proprietary system, splitting its AI strategy at a moment when the company's capital expenditure guidance has hit $115 to $135 billion for the year.** Quick context Mark Zuckerberg is chief executive of Meta, the parent company of Facebook, Instagram, and WhatsApp. Since the release of Llama 2 in 2023, Meta has been the most prominent advocate for open-source frontier AI. The launch of Muse Spark as a closed model on April 8, 2026, marks the first time the company has publicly broken that commitment. ## Llama 5 reaches frontier parity, still free Llama 5, released under an open-weights community licence on April 8, 2026, reaches near-parity with OpenAI's GPT-5 on the two benchmarks most relevant to enterprise developers. On MMLU-Pro reasoning, Llama 5 scores 86.4 against GPT-5's approximately 87; on LiveCodeBench coding tasks, it scores 71.8 percent against GPT-5's approximately 70 percent, a narrow lead. Both figures are from Meta's own benchmarks with Epoch AI's partial external replication, as reported by [FinancialContent/MarketMinute on April 8, 2026](https://www.financialcontent.com/article/marketminute-2026-4-8-meta-unleashes-llama-5-zuckerbergs-open-source-gambit-challenges-proprietary-ai-dominance). The model also supports context windows up to five million tokens, which Meta positions as a practical advantage for document-heavy enterprise workloads where API-gated alternatives impose tighter limits. The competitive point is direct: any developer, startup, or sovereign AI programme can now download a model that rivals the closed frontier on most standard benchmarks at no cost. Meta's internal framing for this approach is the idea of commoditising the complement: making the AI model layer free reduces any single proprietary provider's competitive advantage, while Meta's advertising business and social graph remain insulated from that commoditisation. The company has followed this logic since Llama 2, but Llama 5 is the first version where external benchmark validation makes the parity claim credible rather than aspirational. ![](https://quickchart.io/chart?w=600&h=320&c=%7B%22type%22%3A%22bar%22%2C%22data%22%3A%7B%22labels%22%3A%5B%22MMLU-Pro%20Reasoning%22%2C%22LiveCodeBench%20Coding%22%5D%2C%22datasets%22%3A%5B%7B%22label%22%3A%22Llama%205%20%28open-weights%29%22%2C%22data%22%3A%5B86.4%2C71.8%5D%2C%22backgroundColor%22%3A%22rgba%2824%2C119%2C242%2C0.85%29%22%7D%2C%7B%22label%22%3A%22GPT-5%20%28OpenAI%29%22%2C%22data%22%3A%5B87%2C70%5D%2C%22backgroundColor%22%3A%22rgba%2816%2C163%2C127%2C0.85%29%22%7D%5D%7D%2C%22options%22%3A%7B%22scales%22%3A%7B%22y%22%3A%7B%22min%22%3A60%2C%22max%22%3A95%2C%22title%22%3A%7B%22display%22%3Atrue%2C%22text%22%3A%22Score%20%28%25%29%22%7D%7D%7D%2C%22plugins%22%3A%7B%22title%22%3A%7B%22display%22%3Atrue%2C%22text%22%3A%22Llama%205%20vs%20GPT-5%3A%20Key%20Benchmark%20Scores%22%7D%2C%22legend%22%3A%7B%22position%22%3A%22top%22%7D%7D%7D%7D) Llama 5 vs GPT-5 on two industry benchmarks, April 2026. Source: Meta published benchmarks; Epoch AI partial replication via FinancialContent/MarketMinute, April 8, 2026. ## Muse Spark: Meta's first closed-weights model On the same April morning, Meta Superintelligence Labs (MSL) released Muse Spark, the company's first model with closed weights since the Llama era began. Muse Spark immediately became the engine behind the Meta AI assistant across the company's app ecosystem, with a private API preview extended to select developers, according to [Bloomberg's April 8 coverage](https://www.bloomberg.com/news/articles/2026-04-08/meta-debuts-first-ai-model-from-prized-superintelligence-group). [TechCrunch described the launch](https://techcrunch.com/2026/04/08/meta-debuts-the-muse-spark-model-in-a-ground-up-overhaul-of-its-ai/) as "a ground-up overhaul" of Meta's AI stack. MSL is a division Zuckerberg assembled in summer 2025. He recruited Alexandr Wang, 29, the former co-founder and chief executive of Scale AI, to lead it as Meta's Chief AI Officer. The team rebuilt the full training pipeline from scratch: new model architecture, new data pipelines, new infrastructure. Meta's technical disclosures claim the resulting pretraining stack achieves equivalent capabilities at more than an order of magnitude less compute than Llama 4 Maverick, though that claim remains internally sourced. Muse Spark's initial competitive position is fourth place on the Artificial Analysis Intelligence Index, with a confirmed score of 52, behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6, per [Prism News](https://www.prismnews.com/news/meta-superintelligence-labs-debuts-muse-spark-its-first-ai). The gap to first place is real. Zuckerberg has framed Muse Spark as the beginning of a research programme aimed at what he calls "Personal Superintelligence": an AI agent capable of planning and acting autonomously on behalf of individual users across daily tasks, rather than a finished product. The closed-weights decision reflects that ambition: the company wants a proprietary track at the absolute frontier, separate from the open commoditisation strategy Llama represents. ![](https://quickchart.io/chart?w=600&h=320&c=%7B%22type%22%3A%22bar%22%2C%22data%22%3A%7B%22labels%22%3A%5B%222024%22%2C%222025%20%28est.%29%22%2C%222026%20%28guidance%29%22%5D%2C%22datasets%22%3A%5B%7B%22label%22%3A%22Capital%20Expenditure%20%28%24B%29%22%2C%22data%22%3A%5B38%2C60%2C125%5D%2C%22backgroundColor%22%3A%5B%22rgba%2824%2C119%2C242%2C0.6%29%22%2C%22rgba%2824%2C119%2C242%2C0.75%29%22%2C%22rgba%2824%2C119%2C242%2C1%29%22%5D%7D%5D%7D%2C%22options%22%3A%7B%22scales%22%3A%7B%22y%22%3A%7B%22title%22%3A%7B%22display%22%3Atrue%2C%22text%22%3A%22US%24%20Billion%22%7D%7D%7D%2C%22plugins%22%3A%7B%22title%22%3A%7B%22display%22%3Atrue%2C%22text%22%3A%22Meta%20Annual%20Capital%20Expenditure%20%28%24B%29%22%7D%7D%7D%7D) Meta annual capital expenditure, 2024 actual through 2026 guidance. The 2026 guidance range of $115-135 billion is nearly double 2025 spending. Source: Meta Q1 2026 earnings call via Bloomberg, April 2026. ## One billion users and the Altman mirror Meta AI, the assistant now powered by Muse Spark, crossed one billion monthly active users, per public statements Meta made alongside the April launch. That scale is the commercial foundation for the dual-track strategy: at a billion users, Meta AI is large enough to justify both a free open-source model that expands the developer ecosystem and a proprietary frontier model that defends the assistant's competitive position against OpenAI's ChatGPT and Google's Gemini. The trajectory is instructive when set against Sam Altman's public statements. In a Reddit AMA in January 2025, days after DeepSeek's open-source R1 model rattled AI markets, Altman wrote: "I personally think we have been on the wrong side of history here and need to figure out a different open source strategy." He noted that not everyone at OpenAI agreed with that assessment, per [TechCrunch's January 31, 2025 coverage](https://techcrunch.com/2025/01/31/sam-altman-believes-openai-has-been-on-the-wrong-side-of-history-concerning-open-source/). OpenAI subsequently announced plans to release its first open-weight model since GPT-2. The result is a pair of companies moving in roughly opposite directions at roughly the same time. OpenAI, built on proprietary closure, is edging toward openness. Meta, built on open-source advocacy, is adding a closed proprietary track. Neither has abandoned its prior stance; both now operate across the full spectrum. That convergence complicates any clean narrative about which model of AI development wins. See also: [Sam Altman's portfolio breakdown](https://www.startuphub.ai/ai-news/ai-figures/2026/figure-sam-altman-venture-portfolio-breakdown-2026-05-27) and StartupHub's [earlier coverage of Altman on open source](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/sam-altman-on-ai-s-future-and-open-source). ![](https://quickchart.io/chart?w=600&h=320&c=%7B%22type%22%3A%22bar%22%2C%22data%22%3A%7B%22labels%22%3A%5B%22Q2%202024%22%2C%22Q2%202026%22%5D%2C%22datasets%22%3A%5B%7B%22label%22%3A%22Monthly%20Active%20Users%20%28millions%29%22%2C%22data%22%3A%5B600%2C1000%5D%2C%22backgroundColor%22%3A%5B%22rgba%2824%2C119%2C242%2C0.7%29%22%2C%22rgba%2824%2C119%2C242%2C1%29%22%5D%7D%5D%7D%2C%22options%22%3A%7B%22scales%22%3A%7B%22y%22%3A%7B%22min%22%3A0%2C%22title%22%3A%7B%22display%22%3Atrue%2C%22text%22%3A%22Monthly%20Active%20Users%20%28M%29%22%7D%7D%7D%2C%22plugins%22%3A%7B%22title%22%3A%7B%22display%22%3Atrue%2C%22text%22%3A%22Meta%20AI%20Monthly%20Active%20Users%22%7D%7D%7D%7D) Meta AI monthly active users at two reported milestones. Source: Meta public announcements, 2024 and April 2026. ## What it means Zuckerberg is no longer running a single AI thesis. Llama 5 sustains the ecosystem breadth and developer goodwill that have made Meta a credible AI platform; Muse Spark pursues the frontier capability needed to keep the Meta AI assistant competitive as ChatGPT and Gemini iterate. The $115 to $135 billion capex guidance for 2026, nearly double the prior year, is what funds both tracks simultaneously. Whether the two strategies reinforce each other or eventually pull in opposite directions, as internal model priorities and external open-source commitments potentially conflict, is the central question Zuckerberg will face in the second half of 2026. ## Sources - ![](https://www.google.com/s2/favicons?domain=bloomberg.com&sz=32) [Bloomberg: Meta Debuts First AI Model From New Superintelligence Group](https://www.bloomberg.com/news/articles/2026-04-08/meta-debuts-first-ai-model-from-prized-superintelligence-group) - ![](https://www.google.com/s2/favicons?domain=techcrunch.com&sz=32) [TechCrunch: Meta debuts the Muse Spark model in a ground-up overhaul of its AI](https://techcrunch.com/2026/04/08/meta-debuts-the-muse-spark-model-in-a-ground-up-overhaul-of-its-ai/) - ![](https://www.google.com/s2/favicons?domain=venturebeat.com&sz=32) [VentureBeat: Goodbye, Llama? Meta launches new proprietary AI model Muse Spark](https://venturebeat.com/technology/goodbye-llama-meta-launches-new-proprietary-ai-model-muse-spark-first-since) - ![](https://www.google.com/s2/favicons?domain=financialcontent.com&sz=32) [FinancialContent/MarketMinute: Meta Unleashes Llama 5 -- Zuckerberg's Open-Source Gambit](https://www.financialcontent.com/article/marketminute-2026-4-8-meta-unleashes-llama-5-zuckerbergs-open-source-gambit-challenges-proprietary-ai-dominance) - ![](https://www.google.com/s2/favicons?domain=prismnews.com&sz=32) [Prism News: Meta Superintelligence Labs Debuts Muse Spark](https://www.prismnews.com/news/meta-superintelligence-labs-debuts-muse-spark-its-first-ai) - ![](https://www.google.com/s2/favicons?domain=techcrunch.com&sz=32) [TechCrunch: Sam Altman believes OpenAI has been on the wrong side of history concerning open source](https://techcrunch.com/2025/01/31/sam-altman-believes-openai-has-been-on-the-wrong-side-of-history-concerning-open-source/) Editorial standards: every claim is sourced. Tips: editor@startuphub.ai --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory.