Skip to content

SouthwardTech

Technology, industry and investment across China and the Global South

5 projects tracked1 Contracted4 AnnouncedReviewed

China

China’s hiring giant releases AI tested on tasks from 35 professions

BOSS Zhipin’s Nanbeige4.3-33B targets coding and office work, with developer-reported gains over Qwen on a professional-task benchmark.

BOSS Zhipin, the Chinese recruitment platform, has released Nanbeige4.3-33B, an AI model for coding and office work whose developers report testing it on tasks drawn from 35 professions, from bookkeeping to journalism.

BOSS Zhipin Nanbeige4.3-33B illustration
BOSS Zhipin’s Nanbeige laboratory reports evaluating Nanbeige4.3-33B on JobBench tasks spanning 35 professions. AI-generated illustration: SouthwardTech.

The model comes from the company’s Nanbeige laboratory. Its weights are available on Hugging Face under the Apache 2.0 licence. The team describes the 33-billion-parameter system as a model for both immediate use and training smaller AI agents. It has also published a base checkpoint for developers building general or specialised agents.

The occupational tests come from JobBench, a benchmark built around tasks that professionals say they want to delegate to AI. Its examples include a bookkeeping assignment to reconcile bank statements and accounting records, and a reporter’s task involving conflicting public records and interview notes about drinking-water contamination. Agents work through bundles of files and produce outputs such as spreadsheets and written reports, graded against task-specific criteria.

Nanbeige reports a JobBench score of 44.8, compared with 34.5 for Alibaba’s Qwen3.8-27B in its own evaluation. These are developer-reported scores, not an independent assessment by JobBench’s researchers. They measure performance on selected tasks and cannot establish how much of a person’s job the model could take over.

The same comparison reports higher scores on some coding benchmarks, including 80.9 on Terminal-Bench 2.1 against Qwen’s 73.0. Qwen remained ahead on another coding benchmark, SWE-Bench Pro. The lab says it used the same evaluation settings for each model within each benchmark.

The company behind the release operates a large hiring business. Kanzhun, BOSS Zhipin’s parent, reported average monthly active users of 70.2 million in the second quarter of 2026. Its AI ambitions already extend across the recruitment process.

Kanzhun’s interim report describes tools that help jobseekers refine CVs and practise interviews, alongside AI-assisted candidate searches, chats and interviews for employers. It also discusses earlier Nanbeige models. Those disclosures do not establish that the newly released 4.3-33B powers the recruitment tools.

The lab plans to use the larger model as a teacher for smaller systems, including Nanbeige4.3-3B, transferring capabilities through a process known as distillation. That would make the 33B release a foundation for a wider family of agents. Whether those smaller models can perform useful office work reliably, and at an attractive cost, remains to be demonstrated.

Published 26 October 2026 · Last updated 27 October 2026 · Corrections to corrections@southwardtech.com

Weekly newsletter

The Southward Brief

A weekly briefing on the technologies, companies and industrial shifts reshaping China and the Global South.

Required. Enter a valid email address.

By subscribing, you agree to receive The Southward Brief. You can unsubscribe at any time. Read our Privacy Policy.