The MiroMind team has open-sourced the bAgent model MiroThinker v1.0, offering a 256K context window and the ability to make up to 600 tool calls in a single session. It also introduces for the first time the "Deep Interaction Scaling" framework, advocating for replacing traditional parameter stacking with frequent environmental interactions and real-time feedback to enable agent self-evolution.
The model has integrated a toolchain including search, Linux sandbox, code execution, voice transcription, translation, and more, allowing it to complete complex task loops independently within a few hours. In an official example, MiroThinker used 600 rounds of calls to collect recipes, simulate formulations, calculate calories, iterate on sweetener ratios, and finally output a low-sugar dessert plan with nutritional analysis and cost comparison, all without any human intervention.
MiroMind stated that performance metrics follow the formula "performance ∝ interaction depth × reflection frequency." More tool-feedback cycles can exponentially expand the strategy space. Currently, the model weights and code are available on GitHub and Hugging Face, supporting local deployment with 24GB VRAM, and can integrate with LangChain and LlamaIndex frameworks. Developers can customize their own toolsets to build exclusive evolving agents.
The team revealed that the next step is to expand the tool ecosystem to thousands of calls and explore a "lifelong learning" version with a million-level context. Industry insiders believe that the open-source strategy may trigger an agent arms race, and long-term interaction capabilities may become a key competition field for the next generation of large models.
