Yuankong AI ships Boxer on-device model and two agents, lands HP commercial PC preload

Chinese startup Yuankong AI, founded by a team out of Peking University, has released an on-device model called Boxer along with two agents, Yuankong AI Work for office scenarios and Yuankong AI Science for research, tying model, agent and device into a single execution loop, according to a report by QbitAI.

The commercial landing matters more. According to the report, a few days earlier Yuankong AI's on-device model won a preload deal on HP commercial PCs and the two sides formally signed a strategic partnership. Yuankong AI Work and the Yuankong AI 35B on-device model have been deeply adapted to HP's Z-series ZGX Nano G1n AI workstation, ZBook 8 G2i mobile workstation and Z2 Mini G1a desktop workstation, and are already being packaged into industry solutions for manufacturing, finance, healthcare and research.

The company is not a newcomer. Three years ago the team broke out with ChatExcel, one of China's earlier AI spreadsheet products, and quickly accumulated millions of users; since then its lineup has stretched from an AI DataAgent to an open-source OpenAI4S project positioned against Claude for Science, and to its e2Sci.com research lab, moving from office work into heavier research workflows.

Boxer takes a contrarian stance on size: rather than chasing maximum parameters, it bets on the 20B to 100B mid-size band. The report's reasoning is practical: models under 10B are light enough for chat and simple scheduling, while 200B-plus models are better suited to highly open problems, but high-frequency work such as building reports, handling Excel files or designing experiments has clear boundaries while still testing reasoning, planning and long-chain execution, so mid-size is the better balance between capability and what a device can afford to run.

On capability, Boxer-35B-A3B-v0-0819 scored 73.1 in the company's own WorkArena evaluation, finishing first in its comparison group by 10.3 points, with some tasks beating GPT-4o from more than a year earlier. On deployment, the model still retains 87.3% of its original task performance at 40 TOPS of compute and under 8GB of memory usage.

The agent layer is equally concrete. Yuankong AI Work closes the loop between goals, cross-application execution, result verification, state updates and continued work, so users can state an objective and let the system handle spreadsheets, data lookups, documents, slide decks, code and files. The report stresses that the capability can stay entirely on the local machine, keeping working when the network drops and resuming interrupted complex tasks.

Yuankong AI Science extends that reach from documents and data analysis into real lab equipment, covering literature research, data analysis, bioinformatics and protein engineering while linking experiment design, execution and data回流 back... into a traceable, reproducible pipeline. Underpinning long-running operation is its Memory flywheel, which turns task trajectories, environment events and user feedback into task state, knowledge assets and user preferences, letting agents move from one-shot execution to continuous, cross-task and cross-device work.

The company's thesis is that the growth dividend from static internet data has peaked, while devices sit exactly where tasks happen and keep producing behaviour traces, environment feedback and outcomes. Those traces flow back into post-training, reinforcement learning or memory systems, giving models a path to keep evolving, which is why it compresses the idea into the formula productivity equals on-device AI equals model times agent times device. For Yuankong AI, the deal marks the harder step that edge AI has to clear: moving from technically usable to deployable at scale.

What to watch next is how quickly the partnership extends across HP's product lines. The report says Yuankong AI will follow different HP SKUs into notebooks, workstations and servers and expand overseas through the vendor's global channels, so the real test is whether this model-plus-agent-plus-device combination can deliver reliably in manufacturing, finance, healthcare and research workflows.