The Pulse
IQuest Details Its 320B CLI Model in a New Technical Report
IQuest’s IQuest-Q1 paper describes a 320-billion-parameter model for command-line and coding tasks, alongside benchmark results and deployment guidance. The report says a controlled comparison required about 52% of a Qwen3-30B-A3B-style bas
AI.info Team · Written with AI · fact-checked ·
“We release IQuest-Q1 as an open-source model to support community research and development.”
The IQuest Research paper authors write
IQuest published its IQuest-Q1 technical paper on arXiv on October 22, 2026, and announced the report the following day. The model is a sparse mixture-of-experts system designed for command-line interaction, coding and other multi-step tasks. The paper sets out the training process and benchmark results; the project’s GitHub repository provides deployment instructions.
IQuest’s 320B model targets terminal work
IQuest-Q1 has about 320 billion parameters, with an estimated 15 billion active for each token. Its architecture routes inputs through eight of 256 experts, and the model supports a context window of 524,288 tokens. IQuest says pre-training and mid-training covered approximately 29 trillion tokens.
The paper describes a three-stage pipeline: pre-training, mid-training and post-training. IQuest says the post-training stage combines supervised fine-tuning, domain-specific reinforcement learning and multi-teacher on-policy distillation, with model merging to combine capabilities from different checkpoints. In a controlled comparison using identical training data, IQuest says the model required about 52% of the compute used by a Qwen3-30B-A3B-style mixture-of-experts baseline to reach the same training loss. That figure describes the comparison, not the total cost of training the released 320B model.
Deployment puts the weights in developers’ hands
The repository directs users to download the model weights and serve them with SGLang or vLLM. Its example configurations set tensor parallelism to eight and include parsers for IQuest-Q1’s tool calls and reasoning format. For agent integration, IQuest supplies instructions for Claude Code and Codex CLI, including recommended versions of each.
The repository provides local serving configurations, rather than describing a hosted API. Its CLI integration instructions call for a gateway compatible with the relevant tool-calling interface, while tool execution and reasoning extraction depend on compatible parsers and an agent harness. The model accepts text only, with no native image, audio or video input.
Scores depend on the benchmark version
IQuest reports 84.5 on CyberGym, 64.6 on DeepSWE v1.1 and 57.5 on SWE-Bench Pro. Its Terminal-Bench 2.1 score is 83.2, while the table gives 17.9 on Terminal-Bench 3.0 and 22.7 on Terminal-Bench 4.0. Those results belong to different benchmark versions and should not be read as a direct trend across an unchanged test.
The comparisons also mix publicly reported scores with results IQuest says it produced itself; the paper marks its own test results with asterisks. For agentic coding tasks, the paper says the evaluation uses Claude Code for most benchmarks and mini-SWE-agent for DeepSWE v1.1, with task-specific runtime limits. IQuest’s internal IQuest-CLIBench score is 53.7, but the paper describes that assessment as its own benchmark for CLI interaction and user experience.
IQuest says human review remains necessary
The report cautions that IQuest-Q1 can overlook constraints, repeat failed attempts or leave CLI tasks unfinished. It also warns that generated code and explanations may be wrong and recommends reviewing changes or checking them with appropriate tests. Those caveats matter for a model designed to act through tools and command-line environments, where a plausible answer is not the same as a verified result.
IQuest’s release pairs open weights with integration recipes, but the published serving examples use eight-way tensor parallelism, and reliable tool use depends on compatible infrastructure. The paper’s benchmark table offers a snapshot of performance, not independent confirmation; IQuest’s stated limitations call for supervision during real CLI work.