IQuestLab (iquestlab.github.io) is the open research and showcase portal for IQuest-Coder (LoopCoder) and scientific foundation models developed by IQuestLab. Built around parameter-efficient Loop Architecture and multi-stage Code-Flow training, the platform showcases coding models capable of matching large MoE architectures with reduced High Bandwidth Memory (HBM) and KV-cache overhead—featuring single-GPU consumer deployment (RTX 3090/4090 for Int4), dual Instruct & Thinking tracks, and live browser-based Canvas / simulation benchmarks.

Official WebsiteHugging FaceGitHub

Scaling software engineering foundation models often creates high High Bandwidth Memory (HBM) demands and KV-cache bottlenecks that limit local and on-premise developer deployment. IQuestLab demonstrates how structural parameter sharing (the Loop architecture) combined with temporal repository-change training (“Code-Flow”) enables models to deliver strong reasoning and codebase understanding with lower inference resource footprints.

Core Technical Architecture & Model Matrix

Feature / Dimension	IQuest-Coder / LoopCoder Specification	Core Engineering & Deployment Advantage
Model Architecture	Loop Architecture (Parameter-Sharing Recurrence)	Matches hundred-billion MoE performance with substantially reduced HBM and KV-cache overhead.
Training Pipeline	Multi-Stage Code-Flow Progression	Pre-Train, Annealing, Mid-Train, and Post-Train tracking code evolution over time.
Reasoning Traces	32k Reasoning Traces & Agent Trajectories	Enforces long-horizon context stability and low error propagation.
Dual Post-Training Paths	Thinking Mode & Instruct Mode	Separates deep step-by-step logic chains from low-latency general coding.
Hardware Deployment	Single H20 (Base/Loop) & Single RTX 3090/4090 (Int4)	Run production-grade code models locally without multi-node server clusters.
Open Ecosystem Lineage	IQuest-Coder-V1, UBio-MolFM, UBio-ARCK, TMAS	Broad research spanning code generation, life sciences, and bio-modeling.

Key Capabilities and Interactive Demos

• Loop Architecture Efficiency: By recycling and sharing parameters across internal model loops, the architecture adds only ~5% to training compute while significantly lowering memory requirements, reshaping the compute-performance Pareto frontier.

• Code-Flow & Repository Evolution Training: Instead of training purely on static code snippets, IQuest-Coder ingests Git change flows, multi-file diff histories, and agentic trajectory data to understand how codebases evolve and refactor over time.

• Complex Front-End & Physics Code Synthesis: Demonstrated directly on the portal through zero-dependency, single-file HTML5 Canvas applications generated by the model:

  • Interactive Particle Typography: Dynamic physics text that scatters on mouse approach and springs back.

  • Neon Space Shooter: Arcade Canvas game with turret aiming, boss waves, particle physics, and screen shake.

  • Boids Flocking Simulation: Real-time multi-agent biological flocking with separation, alignment, and cohesion physics.

• Extended Scientific Research Suite (UBio): Complements coding research with biomedical foundation model benchmarks (UBio-ARCK) and molecular modeling suites (UBio-MolFM).