A 30.5B parameter Mixture-of-Experts model for automating long-horizon, deep information-seeking tasks across the web. It requires Python environment configuration and uses external APIs for web search and page reading during inference.
Project overview
The model activates only 3.3B parameters per token and its developers report high performance across multiple agentic search benchmarks.
Project type
AI Search
Use cases
Search & Research
Deployment
Refer to project documentation
License
Apache-2.0
Best for
AI engineers and researchers who need a large language model designed for long-horizon, deep information-seeking tasks and have the coding knowledge required for Python script configuration.
Key capabilities
A 30.5B parameter MoE model designed for long-horizon, deep information-seeking tasks.
Supports ReAct for evaluating intrinsic abilities and IterResearch 'Heavy' mode using test-time scaling for maximum performance.
Capable of parsing and referencing uploaded document files like PDFs and Excel sheets during research queries.
Provides scripts to evaluate the model on various deep research and agentic search benchmarks.
Limitations and risks
Online demos may have latency or fail intermittently.
External tool QPS limits may affect response times during operation.
Operation involves external services, including Serper.dev for web search, Jina.ai for web page reading, and SandboxFusion for a Python interpreter sandbox. A local-only data boundary is not documented.
Getting started
Setup involves medium difficulty: creating a conda environment with Python 3.10.0, installing pip dependencies, copying and configuring a .env file with API keys for external services such as Serper.dev, Jina.ai, DashScope, and SandboxFusion, downloading model weights, and running an inference script using bash.
Alternatives and comparisons
Provides deep, agentic research with proper citations using multiple LLMs and search engines.
Provides a local macOS application with a built-in inference engine for local models.
A framework to train a language model to retrieve on demand, generate, and critique its own generation through self-reflection.