Our Research
TuriX Parallelum - A reliable multi-role framework for complex and long-horizon destop task
Recent breakthroughs in multimodal large language models (MLLMs) — exemplified by GPT‑4o and Gemini 2.5 Pro — have dramatically expanded what desktop agents can perceive, reason about, and execute. Yet empirical studies repeatedly show that conventional GUI‑automation frameworks suffer cascading failures on long‑horizon, multi‑step tasks, causing reliability to plummet as complexity rises.
TuriX Parallelum addresses this gap with a four‑role, multi‑model architecture that orchestrates planning, action, evaluation and supervision in parallel, sustaining high success rates while completing tasks faster than prior systems.
TuriX - The SOTA vision language model for GUI task execution
Recent evaluations show that general‑purpose multimodal LLMs stumble on two fronts that dominate GUI task success — action‑trajectory design and pixel‑level click precision. They struggle to condense multiple low‑level intentions into a single executable step and mis‑ground GUI elements when layouts are dense or stylized, causing error rates to spike on long‑horizon benchmarks.
TuriX, our fine‑tuned vision‑language model, directly targets these pain‑points and delivers state‑of‑the‑art (SOTA) performance on desktop benchmarks by unifying smarter planning with sub‑pixel grounding.
On the public showdown-clicks and private MacClick benchmarks, TuriX attains 64.38% click-location accuracy—an absolute 0.11% winner than the OpenAI CUA model on showdown-clicks, but a 10% relative improvement over the UI-TARS-72B-SFT model, and achieves end-to-end pass@5 task success rates 68% in realworld MacOS laptop task execution.

Pixel-level click accuracy (%). Higher is better. TuriX matches proprietary Ope_x0002_nAI CUA and surpasses all open-source counterparts except Ace-Control (trained expressly for Showdown-Click).

Performance on the 20 easy GUI tasks. Full = UI-tree + Planner + Evaluator + Super_x0002_visor; w/o E&S = no Evaluator/Supervisor; w/o UI = no UI-tree; P-only = Planner only.
