Moonshot AI introduced Kimi K3, a 2.8 trillion-parameter vision-language model available via API with weights promised by July 27, positioning it as the largest known open-weight model to date. Concurrently, Meta launched Muse Spark 1.1, a vision-language model trained for agentic tasks, alongside its first paid Meta Model API access.

  • Kimi K3 utilizes Kimi Delta Attention and Attention Residuals to achieve training efficiency gains of approximately 2.5 times over its predecessor.
  • On Artificial Analysis' Intelligence Index, Kimi K3 scored 57, trailing only Claude Fable 5 (60) and GPT-5.6 Sol (59), while leading all open models.
  • Kimi K3 topped Arena.ai’s Code Arena WebDev leaderboard with a preliminary Elo of 1,679 and led AutomationBench-AA at 53 percent.
  • Meta Muse Spark 1.1 supports up to 1 million input tokens and ties GPT-5.6 Luna and GLM-5.2 on Artificial Analysis' Intelligence Index at low cost.
  • Muse Spark 1.1 tops tool use leaderboards MCP Atlas and JobBench, featuring adjustable reasoning levels and prompt caching.

Kimi K3's performance challenges proprietary dominance by offering near-frontier capabilities with open weights for fine-tuning and study. Meta's entry into the paid API market positions it as a low-cost competitor in the agentic AI space.