Key Points:
- Meta Platforms has repeatedly postponed the launch of its highly anticipated proprietary AI model, Muse Spark, leaving developers without a firm release date.
- The delay has stretched into nearly two months after Meta’s AI chief told developers to expect a public API release “soon.”
- Engineers are reportedly questioning if Muse Spark represents a large enough performance leap over Llama 4 to justify a public rollout.
- The continuous delays raise serious questions about how quickly Meta can monetize its massive $72 billion capital expenditure budget.
The global race to build the most advanced artificial intelligence systems has hit another major twist, exposing deep technical and organizational challenges in Silicon Valley’s executive offices. On Thursday, June 4, 2026, the Wall Street Journal reported that Meta Platforms Inc. has repeatedly delayed the release of its newest, highly anticipated artificial intelligence model to developers. The flagship model, code-named “Muse Spark,” has remained in development for several months as Chief Executive Officer Mark Zuckerberg attempts to position the social media giant as a dominant force in the generative AI race. However, the repeated delays have left external developers without a planned launch date, raising immediate concerns about the company’s competitive standing against aggressive rivals like OpenAI, Google, and Anthropic.
The delay has now stretched into nearly two months since the company’s AI chief told developers on social media to expect an Application Programming Interface (API) release “soon.” For proprietary “closed-source” models like Muse Spark, an API is the only gateway for external developers to access the underlying software. Companies typically release these tools alongside a new model to maximize developer engagement and build a lucrative, recurring subscription revenue stream. By keeping its advanced model locked on its private servers, Meta is delaying monetization of its massive artificial intelligence investments.
According to people familiar with the matter, internal frustration is mounting within Meta’s AI research and engineering divisions. Company engineers are reportedly struggling to meaningfully improve the Muse Spark large language model’s capabilities compared to earlier versions. While these internal developer API trials currently account for only 1.5% of the overall tech segment’s research target, the strategic significance of this model remains immense. Many are openly questioning whether the new model delivers a large enough performance leap over Meta’s existing, open-weight Llama 4 models—such as Llama 4 Scout and Llama 4 Maverick—to justify a public rollout. If the performance gap is incremental at best, releasing the model could damage the company’s reputation for cutting-edge innovation.
Meta designed Muse Spark as a massive, multimodal AI capable of processing both text and image data simultaneously, enabling advanced applications like real-time health insights and automated creator discovery. The model utilizes a sophisticated Mixture-of-Experts (MoE) architecture with up to 400 billion total parameters, allowing it to handle massive context windows of up to 10 million tokens. While early internal benchmarks show that Muse Spark scores close to competitors like GPT-5.4 and Claude on general reasoning tasks, it still lags behind its rivals in complex software coding and advanced logic capabilities, creating a significant performance gap.
The delay of Muse Spark is particularly significant because it represents a major, structural pivot in Meta’s long-term artificial intelligence strategy. Historically, the company has championed an “open-source” philosophy, releasing its Llama models freely to the global developer community to drive rapid adoption and challenge Google’s monopoly. However, Meta is developing Muse Spark as a proprietary, closed-source system. Transitioning to closed models allows Meta to build a wider economic moat, control distribution channels, and charge premium subscription fees to corporate clients. However, this strategy only works if the proprietary model can beat free, open-weight alternatives.
This is not the first time Meta has struggled to keep pace with its agile competitors. Earlier in 2026, reports surfaced that another next-generation text model, code-named “Avocado,” faced similar delays after internal evaluations showed it performed poorly compared to Google’s Gemini 3.0. Sources revealed that Meta’s leadership even discussed temporarily licensing Gemini from Google to power its consumer-facing AI products while its own engineers worked to resolve Avocado’s performance issues. While no formal licensing deals emerged, the discussion proved that even the wealthiest tech firms are finding the frontier of advanced reasoning incredibly difficult to conquer.
These continuous product delays have raised urgent questions on Wall Street regarding the wisdom of Meta’s enormous capital expenditures on AI infrastructure. The social media giant plans to spend up to $72 billion on capital expenditures this year, with the vast majority of that funding going directly toward purchasing advanced Nvidia graphics processors, constructing massive data centers, and recruiting top-tier researchers to serve its 3.5 billion active users. If Meta’s expensive engineering teams cannot deliver competitive, revenue-generating models on schedule, investors may begin to question whether the company is over-allocating capital to an overhyped tech bubble.
Some industry experts believe that Meta’s developmental struggles are a symptom of a much larger, sector-wide trend: a potential “performance plateau” in generative artificial intelligence. For the past four years, developers have assumed that simply scaling up data centers and pouring more compute into training models would automatically yield smarter, more capable systems. However, as the engineering limits of large language models become clear, many researchers are realizing that raw scale is no longer enough. To achieve true human-level reasoning, the industry must develop entirely new algorithmic architectures rather than relying on current data-hungry transformers.
Ultimately, the repeated delays of the Muse Spark API represent a vital crossroads for Meta Platforms’ artificial intelligence ambitions. While a company spokesman told the Wall Street Journal on Wednesday that Meta is currently testing the API with select partners and plans to release it later this month, the lack of a firm, public launch date keeps the tech community highly uncertain. By transitioning to a closed-source, proprietary model, Mark Zuckerberg has set up a high-stakes gamble that requires his engineers to deliver a massive, uncompromised leap in performance. Whether Meta can successfully resolve Muse Spark’s technical limitations and secure a dominant position in the corporate software market will define its economic future, or leave its multi-billion-dollar investments stranded in the shadow of OpenAI.
