NorthCUA-Pro 397B

NorthCUA-Pro 397B is an open-weights, multimodal computer-use agent released by NorthCUA under the Apache-2.0 license. Built on the Qwen3.5 architecture, the model processes user instructions, UI screenshots, and prior interaction history to generate chain-of-thought reasoning alongside precise desktop and browser tool invocations. Rather than acting as a passive multimodal conversationalist, NorthCUA-Pro is explicitly engineered to sit inside an automated execution loop, driving mouse actions, keyboard events, and OS-level operations.

The model marks a significant shift in agentic computing by demonstrating that open-weights systems can match or surpass trillion-parameter proprietary competitors in sustained UI automation. In technical evaluations, NorthCUA-Pro achieved an 88.64% success rate on OSWorld-Verified for desktop tasks and 73.89% on WebArena-Verified for browser automation, outperforming closed enterprise models such as Qwen3.8-Max and Kimi-K3. For infrastructure engineers and AI practitioners, it represents a premier local AI model 397B parameters 2026 release capable of executing real-world digital labor entirely behind your firewall.

Operating this model locally requires an understanding of its execution requirements. NorthCUA-Pro is designed specifically to interface with the official NorthCUA Harness. It operates on a strict normalized_1000 coordinate space, meaning all GUI coordinates emitted by the model are bounded between 0 and 1000 on both axes rather than tied to raw screen resolutions. The runtime harness handles mapping these normalized outputs to physical screen pixels at execution time.

NorthCUA-Pro utilizes a Sparse Mixture of Experts (MoE) architecture totaling 397 billion parameters, with only 17 billion active parameters engaged per forward token pass. This design provides distinct advantages for local inference:

• NorthCUA-Pro 397B MoE efficiency: By activating only 17B parameters during generation, compute throughput and execution latency per step are comparable to mid-sized dense models. However, the full 397B parameter weight matrix must still reside in memory, separating compute requirements from memory capacity.
• Base Architecture: Built on top of Qwen3.5, inheriting its underlying visual encoder and text tokenization layers while adapting attention mechanisms for interleaved image-text agent trajectories.
• Coordinate Precision: Outputs adhere to the normalized_1000 standard. The model internally maps visual features directly to a universal 0 to 1000 grid across both horizontal and vertical axes, eliminating aspect ratio drift when running across heterogeneous display configurations.
• Training Pipeline: Staged learning began with GUI mid-training across 4 million interaction steps and 100,000 trajectories, followed by long-horizon Supervised Fine-Tuning (SFT) over 3,000 verified workflows, and concluded with joint desktop-web Group Relative Policy Optimization (GRPO).

While context length is not explicitly capped in the model weights, running multi-step desktop tasks accumulates high token volumes rapidly due to high-resolution screenshots. The model relies on the NorthCUA Harness to perform visual compaction and trajectory state pruning, preventing KV cache exhaustion during 100-step operations.

NorthCUA-Pro is purpose-built for automated end-to-end task execution. It excels in environments where traditional API-based integrations are unavailable or brittle:

• Visual Grounding and UI Element Parsing: On the NorthCUA-Pro 397B reasoning benchmark for grounding (ScreenSpot-Pro), the model achieves 93.30% accuracy, alongside 90.96% on OSWorld-G. It reliably identifies small UI icons, unlabeled canvas buttons, text fields, and dynamic dropdown menus across dense desktop interfaces.
• Autonomous Desktop Operations: Capable of navigating complex desktop operating systems (Linux, macOS, Windows). Use cases include manipulating enterprise legacy software, operating CAD tools, exporting reports from native desktop clients, and managing file systems.
• Web Workflow Automation: Outperforming standard web scrapers, NorthCUA-Pro handles browser tasks through visual interpretation. It solves multi-step navigation paths, extracts dynamic table data, completes nested checkout flows, and handles authentication dialogues without relying on DOM tree stability.
• Error Recovery and Self-Correction: Because it reviews screenshot observations after every tool execution, the agent can recognize dialog errors, mistyped inputs, or dropped network connections, allowing it to backtrack and attempt alternative paths.

Evaluating NorthCUA-Pro 397B hardware requirements requires treating VRAM capacity as the primary bottleneck, followed by memory bandwidth.

To run NorthCUA-Pro 397B locally, you must allocate sufficient video memory to store 397 billion parameters plus working memory for visual feature tokens and KV cache:

• 16-bit (BF16/FP16): ~800 GB VRAM. Requires an enterprise cluster such as 8x A100 80GB or 8x H100 80GB GPUs.
• 8-bit Quantization (Q8_0): ~420 GB to 440 GB VRAM. Requires at least 6x 80GB enterprise GPUs.
• 4-bit Quantization (Q4_K_M): ~225 GB to 245 GB VRAM. Fits onto 4x A100 80GB, 3x H200, or a multi-GPU consumer workstation.
• 3-bit Quantization (Q3_K_M): ~175 GB to 195 GB VRAM. Feasible on dual 96GB/128GB unified memory configurations, though with risks of coordinate regression.

The best quantization for NorthCUA-Pro 397B is Q4_K_M. In agentic vision models, quantization below 4 bits introduces precision degradation in spatial coordinate regression, causing mouse clicks to miss small UI targets. Q4_K_M retains element-grounding reliability while reducing storage and memory overhead by nearly 70% compared to FP16.

Practitioners wondering how to run 397B model on consumer GPU setups must plan for multi-card rigs. A single consumer card cannot load this model:

• Consumer GPU Workstations: Running Q4_K_M requires at least ten RTX 4090 (24GB) or RTX 3090 (24GB) cards pooled across PCIe lanes (240GB total VRAM). Due to PCIe bandwidth bottlenecks, tensor parallelism across non-NVLink consumer setups will severely impact execution speed.
• Enterprise Workstations (Recommended): The best GPU for NorthCUA-Pro 397B is an SXM-based node featuring 4x or 8x NVIDIA A100 (80GB) or H100 (80GB) interconnected via NVLink. NVLink provides the interconnect speeds necessary to route weights across cards without stalling inference.
• Apple Silicon: Running this model on an Apple Mac Studio (M2/M3/M4 Ultra with 192GB Unified Memory) is constrained. A 192GB configuration can load an aggressive Q3 quantization, but OS overhead and KV cache allocation will likely trigger system swap, degrading performance.

Because only 17B parameters are active per token, NorthCUA-Pro 397B tokens per second are relatively high once memory bandwidth requirements are met:

• NVLink Enterprise Cluster (4x A100 80GB, Q4_K_M): Expect approximately 20 to 35 tokens per second during generation.
• PCIe Multi-GPU or Offloaded System: Expect 1 to 4 tokens per second due to bus latency during expert routing.

For rapid local deployment, you can load quantized GGUF weights via Ollama or llama.cpp, exposing an OpenAI-compatible endpoint. Point the official NorthCUA Harness to this local endpoint, ensuring that image inputs remain uncompressed and the normalized_1000 execution handler is enabled in your agent environment.

1# Basic Ollama model creation from GGUF
2ollama create northcua-pro-397b -f ./Modelfile
3ollama run northcua-pro-397b

Understanding where NorthCUA-Pro 397B sits against alternative architectures helps determine when to deploy it over smaller or general-purpose options.

Qwen2.5-VL 72B is a dense multimodal model that fits on much lighter hardware (2x RTX 4090 using Q4 quantization, or a single 80GB A100). It exhibits strong general visual reasoning and document parsing. However, Qwen2.5-VL is not out-of-the-box optimized for sequential operating system actions. NorthCUA-Pro outclasses it on OSWorld-Verified because of its dedicated GRPO reinforcement training over millions of UI steps, its integrated recovery loops, and native adherence to UI coordinate contracts. Choose Qwen2.5-VL for lightweight inspection; choose NorthCUA-Pro when an agent must execute reliable multi-step desktop workflows autonomously.

Llama 3.1 405B occupies a similar total parameter tier, but it is a dense, text-only model. Running Llama 3.1 405B locally requires comparable VRAM (~230GB in 4-bit) while generating significantly slower due to computing all 405 billion parameters per token. Furthermore, using Llama 3.1 for GUI automation requires pairing it with an external visual encoder and building translation layers. NorthCUA-Pro 397B provides native vision-to-action routing with 17B active parameter speed, making it vastly superior for agentic automation despite their identical memory footprint tiers.