CUA-Master-27B

CUA-Master-27B is an open-weight, multimodal computer-use agent developed jointly by MMLab CUHK, Tsinghua University, HKUST(GZ), and JD Explore Academy. Built on top of Qwen-3.5-27B, the model specializes in end-to-end OS automation, bridging standard graphic user interface (GUI) interactions with command-line interface (CLI) execution. Rather than serving purely as a conversational multimodal LLM, CUA-Master-27B is engineered to perceive desktop screens, reason through multi-step goals, and issue precise operational commands across diverse operating systems.

The model marks a significant step forward for practitioners looking to run autonomous operating system agents on local hardware. On standard evaluations, it achieves an 80.2 score on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench, and 25.0 on Agents' Last Exam. It also demonstrates broad cross-platform versatility, posting 56.2 on WindowsAgentArena, 43.0 on MacOSWorld, and 59.9 on MacAgentBench. These metrics place it in competition with frontier proprietary computer-use APIs while remaining deployable within self-hosted, sandboxed local infrastructure.

For engineers seeking a dedicated local AI model with 27B parameters, CUA-Master-27B eliminates reliance on cloud-hosted vision APIs for complex UI automation. Its open-weight release on Hugging Face provides developers with full control over execution latency, data privacy, and agentic trajectory rollouts.

CUA-Master-27B features a dense 27-billion-parameter architecture initialized from Qwen-3.5-27B. Because it is a dense network rather than a Mixture of Experts (MoE), every parameter is active during inference. This requires consistent memory bandwidth and compute across each step, but it avoids the routing overhead and complex memory-distribution challenges typical of sparse models.

The model processes text and high-resolution visual inputs jointly. The training regimen relies on a nested double-loop recursive improvement system:

• Iterative Supervised Fine-Tuning (SFT): Trained on roughly 115,000 tasks stacked across three iterations (63K, 38K, and 14K tasks). These seed trajectories are mined from real-world workflows, including instructional desktop tutorials, and converted into compositional primitives via visual-semantic trajectory reconstruction.
• Reinforcement Learning (RL): Refined using approximately 12,000 high-quality verification tasks. A feedback-guided snapshot-rollback retry mechanism was used during data collection to build robust recovery paths for failed interactions.

While the exact context length is not formally specified in the base documentation, handling agentic GUI trajectories demands sufficient sequence budget to track historical screenshot embeddings, tool definitions, shell outputs, and prior interaction attempts. When deploying CUA-Master-27B, the dense visual encoder translates desktop captures into visual tokens that feed directly into the core transformer alongside standard action schemas.

CUA-Master-27B is built specifically around instruction-following, visual reasoning, and code generation within execution environments. It does not treat desktop tasks merely as mouse-click coordinate prediction; it actively switches between visual pointing and bash-level automation depending on task efficiency.

Key operational capabilities include:

• Hybrid GUI and CLI Workflows: The agent can navigate a browser or desktop application to download an asset, inspect the file structure, switch to a terminal to manipulate the data using bash or Python scripts, and verify the output visually.
• Visual Spatial Reasoning: Parsing complex UI components, nested application menus, high-density system dashboards, and dynamic web layouts. It reads UI states to verify if previous actions took effect before executing the next sequence.
• Self-Correction and Error Recovery: Leveraging its RL training, the agent recognizes OS error dialogs, failed terminal exit codes, or misclicks, rolling back its plan rather than falling into unrecoverable interaction loops.
• Cross-Platform OS Automation: Capable of running workflows across Linux, Windows, and macOS, adapting seamlessly to differences in system menus, shortcut keys, and file system navigation.

To run CUA-Master-27B safely, always deploy the agent in a sandboxed virtual machine or isolated container environment. Because the model issues raw system commands and GUI clicks, direct execution on bare-metal host systems presents substantial security and data integrity risks.

Successfully hosting this 27B model locally requires planning for both weight storage and the dynamic KV cache overhead that builds up as sequential desktop screenshots are ingested.

The table below breaks down the minimum and recommended VRAM allocations across common execution formats:

For most practitioners, the best quantization for CUA-Master-27B is Q4_K_M. This retains visual token fidelity and spatial reasoning capabilities while reducing the core weight footprint to approximately 17 GB.

• Single Consumer GPU (24GB VRAM): Running Q4_K_M on an Nvidia RTX 3090 or RTX 4090 allows the model weights to fit comfortably into VRAM. However, multi-turn UI interaction buffers will quickly consume the remaining 7 GB of headroom as screenshots accumulate. For extended agent runs, context pruning or a slightly more aggressive quantization (like Q3_K_M) may be required to prevent out-of-memory errors.
• Dual Consumer GPUs (48GB VRAM): Two RTX 3090 or 4090 cards linked via PCIe/NVLink represent the best GPU configuration for CUA-Master-27B at home or in the lab. This setup comfortably runs Q8_0 or uncompressed FP16 quantization with ample headroom for full context trajectories and fast KV caching.
• Unified Memory Systems: Mac workstations equipped with an Apple Silicon M3/M4 Max or M2/M3 Ultra (64 GB unified memory or higher) run 4-bit and 8-bit quantized variants with ease, bypassing GPU memory partitioning constraints entirely.

On an Nvidia RTX 4090 using optimized runtimes (such as vLLM or llama.cpp), expect the following CUA-Master-27B tokens per second performance:

• Q4_K_M: 30 to 45 tokens per second for pure text generation. Time-to-first-token will vary depending on the resolution and number of screen images processed in the prompt.
• Q8_0 / Dual GPU: 35 to 50 tokens per second across split layers.

For quick deployment, you can run CUA-Master-27B locally via Ollama once the GGUF weights are converted or indexed in the library:

1ollama run cua-master:27b

For production agent setups, deploy via vLLM or SGLang with OpenAI-compatible visual API endpoints configured, connecting the model output to an agent harness like OSWorld or custom pyppeteer/pyautogui execution scripts.

Evaluating CUA-Master-27B performance against similar architectures highlights its specialized focus.

UI-Mate-27B is its direct architectural contemporary in the 27B parameter class. In direct benchmark evaluations, CUA-Master-27B demonstrates substantial leads:

• Outperforms UI-Mate-27B by 2.6 points on OSWorld Verified (80.2 vs. 77.6).
• Extends the lead on OSWorld 2.0 with a 12.9-point margin (36.9 vs. 24.0).
• Dominates complex workflows, showing a 25.6-point lead on WeaveBench (43.9 vs. 18.3) and an 11.8-point lead on Agents' Last Exam (25.0 vs. 13.2).

UI-Mate remains capable for straightforward web UI clicking, but CUA-Master-27B is far superior when a task requires pivoting between GUI actions and shell scripting.

While the underlying Qwen-3.5-27B foundation model is a strong general-purpose reasoning and code engine, it lacks the interaction grounding necessary to control desktop environments reliably. Standard instruct models often hallucinate non-existent screen coordinates, misunderstand UI bounding boxes, and lack recovery loops when an executed bash command returns an exit code of 1. CUA-Master-27B sacrifices marginal open-ended conversational flair in exchange for hardened, production-grade OS trajectory navigation.