The open-source AI community has released MiroThinker 1.5, a deep research agent from the MiroMind AI team. Unlike models that scale parameters, it uses Interactive Scaling—improving through repeated environment interaction, trial-and-error, and reflection. It's available in 30B and 235B parameter sizes, built for long-horizon tasks.

MiroThinker employs a 'Reasoning-Verification-Correction' loop, processing a 256K context window and supporting up to 400 tool calls per task. The 30B model outperforms Kimi-K2-Thinking on a Chinese web benchmark at 1/20th the inference cost. The 235B version scored 80.8% on GAIA-Val-165, a top-tier general AI assistant benchmark.

Its training enforces strict causality to prevent 'spoiler hallucinations,' making it viable for financial prediction and market trend analysis. The team claims it successfully predicted A-share limit-ups and GTA 6 release trends. This focus on 'slow thinking' addresses a key gap in current LLMs, which often hallucinate answers instead of verifying facts.