IQuest Coder
IQuest-Coder-V1 is IQuestLab's family of coding language models for software-engineering research and code assistance. The repository documents model deployment and evaluation rather than a hosted subscription app. Developers who want to experiment with model behavior or build a coding agent can inspect its model variants and implementation guidance.
Choose a model for the task
The family includes 7B, 14B, and 40B models with Instruct and Thinking variants. IQuestLab describes Instruct models as suited to general coding assistance and instruction following, while Thinking models produce explicit reasoning for more complex problems. All models support a context window of up to 128K tokens, according to the repository.
The 40B Loop variants use a recurrent Transformer design with shared parameters across two iterations. The publisher identifies 40B-Loop-Thinking as an experimental research prototype for studying loop-based computation alongside explicit reasoning. Its purpose should guide an evaluation; a research architecture does not establish better results for your application's workload.
Deployment requires a configured environment
The README provides a Transformers quickstart and vLLM serving instructions. IQuest-Coder-V1 uses custom modeling code through Hugging Face's auto_map feature. The publisher recommends Transformers version 4.52.4 or newer.
The CLI-agent section lists supported variants and shows how to connect a served model to Claude Code. Tool-parser settings vary by model, and the instructions warn against setting a reasoning parser for Instruct models. Read those settings for the exact variant you choose before treating it as an interchangeable backend. The repository also describes preliminary HTML and SVG generation support.
Evaluate generated code and sustained agent work
The models generate code but do not execute it. Review outputs and validate them in a sandbox before relying on an implementation. IQuestLab warns that plausible code can be incorrect and that performance can vary for specialized or proprietary frameworks.
The repository also acknowledges weaker instruction following in some situations and limited performance on long-horizon tasks or multi-turn tool use. A large context window therefore needs to be assessed separately from sustained context management. IQuestLab supplies a SWE-Bench Verified evaluation framework and trajectory data for reproduction; its advertised benchmark results are publisher reports, not evidence that a particular deployment will match them.