AI Papers Reader

Personalized digests of latest AI research

View on GitHub

IQuest-Q1 Aims to Make Open AI Models Better at Command-Line Work

IQuest Research has released IQuest-Q1, an open-weight AI model built to operate command-line interfaces (CLIs), the text-based environments where software tools, files, and programs are controlled through typed commands. As language models shift from answering questions toward carrying out multistep tasks such as fixing code or managing systems, their reliability in these environments has become a practical concern. The authors argue that more capable agents require stronger CLI skills.

IQuest-Q1 uses a sparse Mixture-of-Experts (MoE) design, in which many specialized sub-networks exist but only a few activate for each token. The model has 320 billion total parameters, of which 15 billion are active at once. Its training has three main stages. Pretraining uses a two-stage data curriculum that raises the share of code in the training mix from 40% to 70%. Mid-training extends the context window, the amount of text the model can consider at once, up to 512,000 tokens. Post-training combines supervised fine-tuning, reinforcement learning in task environments, and multi-teacher on-policy distillation, in which a student model learns from several frozen expert models. The team then merges parameter updates from different experts and checkpoints, which the authors describe as a way to combine skills while limiting conflicts between them.

The paper illustrates a failure mode with a long-horizon task in which the model made 35 tool-call errors because it mismatched tool names and parameters. An agent built on IQuest-Q1 traced 6,054 tool calls and found that the training data was heavily biased toward one tool-call format, with some records containing mismatched definitions. Human experts reviewed the diagnosis and approved a retraining run on 2.35 billion tokens, consuming 938.7 GPU hours. The authors say the new checkpoint was tested on both familiar and held-out schemas, but the paper does not report the outcome.

On the coding benchmark DeepSWE v1.1, IQuest-Q1 scores 64.6, ahead of GLM-5.3 Flash (63.4) and DeepSeek-V4-Flash-0731 (54.4), but behind Claude Opus-5 (73.7) and GPT-5.6 Sol (73.0). On CyberGym, which tests reproduction of real vulnerabilities, it scores 84.5 against 81.6 for GLM-5.3 Flash and 83.6 for GPT-5.6 Sol. On Terminal-Bench 2.1 it scores 83.2, below Claude Opus-5 (89.1), GPT-5.6 Sol (88.8), and GLM-5.3 Flash (84.3). The gaps widen on Terminal-Bench 3.0 (17.9, versus 42.7 for Claude Opus-5) and 4.0 (22.7, versus 53.9). On the authors’ in-house IQuest-CLIBench, the score is 53.7, compared with 58.2 for Claude Opus-5 and 58.5 for GPT-5.6 Sol. In a controlled comparison with a Qwen3-30B-A3B-style baseline, the authors report 1.93 times the scaling efficiency, meaning roughly 52% of the training compute reaches the same loss.

These results have caveats. Benchmark setups differ, with several run under different harnesses and time limits, and IQuest-CLIBench is proprietary. The authors’ stability projections, including a critical model size of about 572 billion active parameters, are extrapolations from fitted trends. The paper also lists known weaknesses: the model may overlook constraints, repeat failed attempts, and need human review.

The work’s significance lies mainly in its openness and its design. IQuest-Q1 is released with weights and deployment details, and its training pipeline is built around CLI interaction. The authors describe a human-supervised loop in which the model helps build its successors, a step they frame as a pathway toward recursive self-improvement. That claim rests on a single illustrative iteration, so it remains an early indication rather than an established result.