Open-weights computer-use model family (9B, 35B, 397B MoE) from Beipo AI, built on Qwen3.5, with an open-source harness.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
NorthCUA is a family of computer-use models from Beipo AI, released in three sizes: Nano (9B dense), Mini (35B total, 3B active MoE) and Pro (397B total, 17B active MoE). The models are built on Qwen3.5 and take user requests, screenshots and interaction history, then produce reasoning, responses and tool calls for a desktop or browser runtime.
The model card states that the models must be run with the official NorthCUA Harness, which handles prompts, tool calls, observations and action execution. GUI action coordinates use a 0–1000 normalized space rather than raw pixels, and the harness converts them to pixels at execution time.
Results reported in the technical report (not new measurements of the released files, according to the card) give OSWorld-Verified success of 78.90% (Nano), 85.30% (Mini) and 88.64% (Pro), and WebArena-Verified of 67.16%, 70.43% and about 73.9%. Grounding scores for Pro include 93.30% on ScreenSpot-Pro and 90.96% on OSWorld-G. A separate OSWorld 2.0 figure of 66.80% mean partial credit for Pro is not a full-completion rate. Weights are on Hugging Face under Apache-2.0; the full 397B Pro model needs substantial memory despite the 17B active parameters.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open-weights computer-use model family: Pro (397B total/17B active MoE), Mini (35B) and Nano (9B), with an open-source harness. It scores 78.9/85/88 on OS World Verified for Nano/Mini/Pro.
Closest neighbours by meaning, not just by category.
Model · Accio Lab
Open-weights 35B MoE (3B active) agentic model from Accio Lab, post-trained from Qwen3.6-35B-A3B.
Model · Meituan
Meituan's open-weights 1.6T-parameter multimodal MoE model (48B active) for agentic coding and visual tasks.
Model
35B Mixture-of-Experts open-weights model from InternScience for long-horizon agentic search, engineering and tool use.
Model · Nex AGI
Open agentic model family from Nex AGI for computer use, coding and long-horizon tool tasks, in several sizes.
Model · inclusionAI
Open-weights 560B hybrid-linear MoE reasoning model (25B active, 1M context) from inclusionAI, MIT licensed.
Model
Open model family (9B dense, 35B MoE, 397B MoE) trained with a self-improvement loop for reasoning, coding and agentic tasks.
