Meta has unveiled Muse Spark 1.1, an agentic AI model with a one-million-token context window, in its most aggressive bid yet to challenge OpenAI and Google at the frontier of artificial intelligence.
Released around July 9, 2026, Muse Spark 1.1 is positioned as a direct competitor to OpenAI's GPT-5.6 Sol and Google's Gemini 3.6 Flash. The model features computer-use capabilities, subagent orchestration, and integration with Meta's existing platforms, including Replit for developer workflows. Unlike previous Meta AI releases, Spark 1.1 does not ship with open weights, a shift that reflects the soaring cost of training frontier models.
The announcement came alongside a broader infrastructure push. Meta confirmed plans to begin manufacturing its Iris AI chip in September, aiming to double its computing capacity to 14 gigawatts by next year. The company is also building a cloud business, Meta Compute, to sell excess AI infrastructure capacity to external customers, marking a significant diversification beyond advertising.
Muse Spark 1.1's agentic benchmarks have drawn mixed but generally positive early reviews. On the MCP Atlas benchmark, which measures multi-step task completion, Spark 1.1 reportedly rivals top-tier models from OpenAI and Anthropic. Meta is offering the model through a new paid developer API, its first foray into charging third parties for direct model access.
The release was not without controversy. Meta simultaneously launched Muse Image, a generative image tool integrated into Instagram Stories and WhatsApp, but pulled it back within days after privacy advocates raised concerns about automatic use of public Instagram photos. The incident highlighted the regulatory and reputational risks Meta faces as it integrates AI more deeply into its social platforms.
Mark Zuckerberg has publicly emphasized openness in AI development, but the decision to keep Spark 1.1 closed suggests a pragmatic pivot toward protecting competitive advantage. With AI capital expenditure expected to exceed $100 billion industry-wide in 2026, Meta is betting that owning the full stack, from chips to models to cloud services, will yield better returns than open-sourcing its most valuable research.
Image source: i.ibb.co