NorthCUA-Pro 397B Beats Trillion-Param Closed Models

An open-source computer-use agent has achieved a benchmark milestone: NorthCUA-Pro, a 397-billion-parameter model built on Qwen3.5, outperforms trillion-parameter closed systems Qwen3.8-Max and Kimi-K3 on verified desktop and web benchmarks. The technical report, posted to Hugging Face Papers, documents a three-part architecture and training pipeline that pairs long-horizon reinforcement learning with 4 million GUI interaction steps collected across dozens of applications.

Release Date and Availability

NorthCUA officially launched on November 3, 2026, with full open-source access to three model sizes: Nano at 9 billion parameters, Mini at 35 billion, and Pro at 397 billion. The release includes the Harness execution framework, trained model weights, and documentation covering the Experience Flywheel data collection system and Staged Learning pipeline. All trajectories were collected in isolated E2B sandboxes and filtered by task outcome.

Benchmark Performance Against Closed Models

On OSWorld-Verified, the desktop benchmark suite, NorthCUA-Pro surpasses both Qwen3.8-Max and Kimi-K3, while Mini beats Kimi-K3. On WebArena-Verified, the web interaction benchmark, Mini outperforms Qwen3.8-Max and Pro beats both closed competitors. The report frames these as comparative verdicts rather than publishing absolute scores, positioning the 397B open model ahead of systems with roughly 2.5x more parameters.

Architecture and Training Pipeline

The NorthCUA system integrates three co-designed components. The Experience Flywheel supplies executable tasks drawn from real-world application usage. The Harness sustains long interactions, records agent decisions, and validates outcomes across extended sessions. Staged Learning moves from grounding mid-training through long-horizon supervised fine-tuning to joint desktop and web reinforcement learning using Group-Relative Policy Optimization. Training scale reached over 4 million GUI interaction steps, with data collected across desktop applications and web environments.

Speed-Optimized Ultrafast Variant

A tradeoff variant called NorthCUA-Ultrafast pairs direct action selection with generation on the pre-RL Nano model. The report credits this configuration with 2.3x faster average completion time on matched successful WebArena-Verified tasks compared to the standard inference path. The speed gain comes from bypassing intermediate reasoning steps in favor of learned action mappings, a design choice suited to latency-sensitive deployments where task completion speed matters more than maximum success rate.

What This Means

NorthCUA-Pro's performance demonstrates that open-source computer-use agents can match or exceed closed trillion-parameter systems at less than half the parameter count when training architecture, data quality, and reinforcement learning are co-optimized. The release arrives into a dense cluster of agent framework work, with over 527 agent-related stories tracked in the past 90 days. By shipping three model sizes, full training code, and the execution harness, the project lowers the barrier for researchers and developers building GUI automation systems on verified desktop and web benchmarks.