Chen Tianqiao’s MiroMind Touts “Top-Tier” Predictive AI Model After Back-to-Back Wins on FutureX Benchmark
Want to read in a language you're more familiar with?
MiroMind says its memory-driven predictive AI topped the FutureX benchmark for two straight weeks—shifting focus from text generation to real-time prediction.
MiroMind, the AI startup founded by Chinese entrepreneur and philanthropist Chen Tianqiao, says it has built a world-class predictive large model—distinct from text-generation LLMs—after its agent framework MiroFlow topped the FutureX real-time forecasting benchmark for two consecutive weeks in September. Company materials describe MiroMind’s approach as memory-driven and purpose-built for prediction and decision-making rather than pure text output.
According to multiple Chinese tech outlets summarizing the FutureX leaderboard, the latest MiroFlow runs used GPT-5 as the base model and also ranked highly when paired with MiroMind’s own MiroThinker model. The team publicized case studies that included correctly forecasting ATP men’s singles rankings and calling key price levels for Solana (SOL) during a September window, to showcase systematic reasoning and risk modeling.
MiroMind has been developing both agent frameworks and foundation models this year. In August, the team—co-led by Tsinghua University associate professor Dai Jifeng—released MiroMind Open Deep Research (Miro ODR) as a fully open-source “deep research” stack. The initial version posted an 82.4 score on GAIA, outperforming several open- and closed-source peers cited in press materials, and the group says it plans monthly open-source updates.
MiroMind describes itself as a globally distributed AI company focused on foundational AGI research and “next-generation” predictive intelligence. Chen—best known as the founder of Shanda and the Tianqiao & Chrissy Chen Institute for Neuroscience—has positioned the venture as an open, research-heavy effort with ambitions to rival leading labs.
While most headline LLMs optimize for generation (chat, coding, content), real-time prediction under uncertainty is a different problem class. If FutureX results generalize, MiroMind’s memory-driven architecture could expand how agentic systems handle markets, sports, logistics and other fast-changing domains. Independent replication of the FutureX evaluations and the real-world trading/operations efficacy will be the next tests.
