BUZZ BEAT MEDIA

Source: PR Newswire · IQuest Research · View original release

BEIJING, Oct. 23, 2026 /PRNewswire/ — IQuest Research has published IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems, detailing the model’s architecture, complete training pipeline, results across 30 benchmarks, and role in iterative model development.

IQuest-Q1 is an open-source agentic foundation model optimized for command-line interface (CLI) interaction, software engineering, and complex multi-step tasks. Its sparse Mixture-of-Experts architecture comprises 320 billion total parameters, with 15 billion activated per token.

A Complete Training Pipeline, Beginning with Pre-Training

Developed from scratch, IQuest-Q1 progresses through pre-training, mid-training, and post-training. Pre-training and mid-training together cover approximately 29 trillion tokens, with context length progressively extended to 512K.

The architecture incorporates Stable SplitNorm, the Muon optimizer, and μP parameterization. In controlled experiments using identical training data, it required approximately 52% of the compute needed by a Qwen3-30B-A3B-style MoE baseline to reach the same training loss, corresponding to 1.93× scaling efficiency. The report also introduces Critical Model Size (CMS) to examine scaling limits under training-stability and optimization-efficiency constraints.

Post-training combines supervised fine-tuning, domain-specific reinforcement learning, Multi-Teacher On-Policy Distillation (MOPD), and model merging. Training spans multiple agent harnesses and versions, including Claude Code and Codex, to strengthen performance across different execution environments.

The report also describes Recursive Multi-Token Prediction, which adapts an MTP layer into a reusable drafter for speculative decoding, extending the drafting horizon within a fixed parameter budget.

Evaluations Across Coding and Agentic Workflows

The 30-benchmark evaluation suite covers software engineering, terminal operation, tool use, professional workflows, search, and reasoning. IQuest-Q1 scores 52.7 on AutomationBench, 51.8 on τ³-Banking, 74.1 on SkillsBench, 78.4 on MCP-Atlas, and 75.0 on Toolathlon-Verified.

The report also evaluates sustained interaction through Claw-Eval and IQuest-CLIBench, the team’s internal assessment combining user simulation, execution traces, and environment-state checks.

Full results, comparison baselines, and evaluation configurations are disclosed in the report.

Environment Integrity and Reward Hacking Mitigation

The team constructed more than one million executable, containerized task instances, retaining valid and complete trajectories for mid-training. Post-training environments use task-specific verifiers to ground learning in execution outcomes.

Environment preparation, execution-time checks, and trajectory auditing help mitigate reward hacking, including attempts to access hidden tests or leaked solutions. Confirmed cheating trajectories are excluded from training.

The report also documents model-assisted research under human supervision. In one experiment, an IQuest-Q1-based agent traced 6,054 tool calls to investigate 35 errors, identified training-data inconsistencies, and implemented a reviewed data-cleaning and augmentation plan. Validated methods and research assets will support subsequent model development.

Project resources are available through:

• Technical Report: https://arxiv.org/pdf/2610.30949
• Blog: iquestlab.github.io
• Hugging Face: IQuestLab/IQuest-Q1
• GitHub: IQuestLab/IQuest-Q1

CONTACT: IQuest Research
[email protected]