Meta pivots to Muse Spark
Meta is reportedly moving away from its open Llama model line toward a proprietary model called “Muse Spark,” which enterprise users in Southeast Asia say raises API-cost and data‑control concerns. The reporting frames this as a shift from openness to tighter commercial control over models, a change discussed in recent analysis. (dev.to)
Meta has begun shifting attention from its open Llama line to a new proprietary model, Muse Spark, opening it first inside Meta products and in a private API preview for selected partners. (about.fb.com) Meta announced Muse Spark on April 8, 2026, calling it the first model in a new Muse series built by Meta Superintelligence Labs after “the last nine months” of rebuilding its AI stack. The company said Muse Spark already powers the Meta AI app and website and will roll out to WhatsApp, Instagram, Facebook, Messenger, and Meta’s AI glasses in the coming weeks. (about.fb.com) Meta did not announce a broad public enterprise API at launch. Its April 8 post said Muse Spark would be offered “in private preview via API to select partners,” while the company described the model as “purpose-built for Meta’s products.” (about.fb.com) That marks a visible change in emphasis from Meta’s Llama campaign. On July 23, 2024, Mark Zuckerberg wrote that “Meta is committed to open source AI” and said future Llama models would become “the most advanced in the industry.” (about.fb.com) Meta reinforced that message on March 18, 2025, when it said Llama had passed 1 billion downloads and called open sourcing AI “essential” to widening access. The company said enterprises, startups, universities, nonprofits, and public institutions were already building on Llama. (about.fb.com) The cost issue sits at the center of the backlash from some enterprise users. Meta itself promoted open models in May 2025 by citing Linux Foundation research that found two-thirds of surveyed organizations believed open-source AI was cheaper to deploy than proprietary models, and nearly half cited cost savings as a reason for choosing it. (about.fb.com) A new analysis from Thailand argues that companies which fine-tuned and self-hosted Llama now face a different budget model if they must buy access through a proprietary API. The article says some Thai firms used local Llama deployments for customer-service systems, retrieval tools, and review analysis, and warns that per-token charges could turn fixed infrastructure spending into variable monthly bills. (dev.to) That same analysis ties the shift to data-control worries in Southeast Asia, especially for companies that kept models on their own servers to limit where sensitive business or customer information traveled. The piece frames the problem around Thailand’s Personal Data Protection Act, though it does not cite a direct Meta policy requiring customers to move regulated data into Muse Spark. (dev.to) Meta’s public line is that Muse Spark expands what Meta AI can do, including multimodal input and parallel “subagents” that split up tasks such as trip planning. The company has not, in the materials reviewed here, said Llama is discontinued, but its newest flagship announcement puts the commercial spotlight on a closed model inside Meta’s own ecosystem. (about.fb.com) For companies that chose Llama because they could download it, fine-tune it, and run it themselves, the question is no longer whether Meta still talks about openness. It is whether Meta’s next wave of top-end AI will remain something customers can own, or mainly something they rent. (about.fb.com)
