#China #Beijing #IQuest-Q1 #IQuest Research #CLI Systems

IQuest Research Launches Comprehensive IQuest-Q1 Technical Report

IQuest Research has recently released its technical report on IQuest-Q1, officially titled Advancing Foundation Capabilities for Agentic CLI Systems. This report provides an in-depth look at the model's architecture, training pipeline, and performance across various benchmarks that highlight its capabilities in command-line interaction and software engineering.

Model Overview

IQuest-Q1 stands out as an open-source agentic foundation model designed specifically for command-line interface interactions, as well as for tackling complex, multi-step tasks. At the core of its architecture is a sparse Mixture-of-Experts (MoE) system, which comprises a staggering 320 billion parameters, activating 15 billion of these per token, thus allowing it to handle intensive computational tasks efficiently.

Training Pipeline Breakdown

The development of IQuest-Q1 commenced from scratch, progressing systematically through distinct phases: pre-training, mid-training, and post-training. Collectively, the pre-training and mid-training phases cover an impressive 29 trillion tokens, with the model's context length progressively extending to 512K tokens.

Among the sophisticated optimizations included in the architecture are Stable SplitNorm, the Muon optimizer, and μP parameterization. Notably, the training efficiency of IQuest-Q1 is remarkable; it has demonstrated approximately 52% of the compute requirements compared to a Qwen3-30B-A3B-style MoE baseline, achieving the same training loss with a 1.93× scaling efficiency. This efficiency is crucial for managing resources while maintaining high performance.

The report further introduces a new metric, Critical Model Size (CMS), which aids in evaluating the scaling limits of the model under various training stability and optimization efficiency constraints.

Post-Training Enhancements

After completing the core training, the process includes several enhancements such as supervised fine-tuning, domain-specific reinforcement learning, and Multi-Teacher On-Policy Distillation (MOPD). These processes utilize various agent harnesses and versions—including Claude Code and Codex—to optimize performance in different execution contexts, ensuring robustness across platforms.

Moreover, the report details the Recursive Multi-Token Prediction method, an innovative approach that integrates a Multi-Token Prediction (MTP) layer into a reusable drafting format. This method significantly extends the predictive capabilities of IQuest-Q1 within fixed parameter budgets.

Benchmark Evaluation Results

A highly comprehensive evaluation suite consisting of 30 benchmarks effectively assesses IQuest-Q1’s performance across several domains such as software engineering, terminal operations, tool usage, and reasoning tasks. The results showcase

• - 52.7 on AutomationBench
• - 51.8 on τ³-Banking
• - 74.1 on SkillsBench
• - 78.4 on MCP-Atlas
• - 75.0 on Toolathlon-Verified

Furthermore, sustained interaction assessments through Claw-Eval and the internal IQuest-CLIBench also emphasize the model’s efficiency through user simulation and execution evaluations.

Ensuring Environment Integrity

To further enhance the training process, IQuest Research has constructed over one million executable, containerized task instances. Each instance retains valid trajectories for mid-training, and specific task verifiers are used during the post-training phase to secure learning grounded in actual execution outcomes. The model's preparations and execution checks work collectively to hinder reward hacking attempts, thereby maintaining integrity in the training process. Any trajectory found compromised is excluded from further training iterations, upholding data quality and reliability.

Scientific Contributions and Future Prospects

The report also highlights the successful model-assisted research efforts under human supervision, where IQuest-Q1 played a pivotal role in tracing over 6,054 tool calls to help resolve 35 errors related to inconsistencies in training data. This instance led to a carefully vetted data-cleaning and augmentation plan.

The validated methodologies and research assets developed throughout this project will undoubtedly pave the way for future advancements in foundation models like IQuest-Q1.

For those eager to dive deeper into the findings of this technical report and to access the associated resources, they are available through the following links:

• - Technical Report
• - Blog
• - GitHub Repository

This release marks a significant milestone for IQuest Research as it prepares to shape the future of agentic systems and command-line interfaces through innovative technology.