---
id: "2026-10-24-nanbeige4-3-33b-open-weights"
url: "https://postcutoff.com/e/2026-10-24-nanbeige4-3-33b-open-weights/"
as_of: "2026-11-13T13:49:00+01:00"
date: "2026-10-24"
date_precision: day
category: open-source
importance: 2
confidence: high
status: [Confirmed]
sources: 2
editor: human editor
human_review: null
version: "2026-10-27"
---

As of: 2026-11-13 13:49 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-24-nanbeige4-3-33b-open-weights/

# BOSS Zhipin's Nanbeige team open-sources Nanbeige4.3-33B, a 512K-context agentic coding model under Apache 2.0

On Oct 24, 2026 the Nanbeige team at Kanzhun (operator of China's BOSS Zhipin recruitment platform) published Nanbeige4.3-33B and its base model on Hugging Face under Apache 2.0. The ~33B model has a 512K-token context and is tuned for coding agents and office tasks; the team reports it beats Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B on most agentic benchmarks it ran, figures not yet checked independently. It is meant as a teacher for distilling agent skills into smaller models such as a planned 3B.

## Key facts
- Hugging Face repos Nanbeige/Nanbeige4.3-33B and -Base created Oct 24, 2026; Apache 2.0 (commercial use allowed); ~33.1B parameters
- Context window 524,288 tokens (512K); enable_thinking / preserve_thinking switches; XML or JSON tool calls
- Team comparison against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at highest thinking effort with the same harness: Terminal-Bench 2.1 and 4.0, SWE-Atlas-QnA, NL2Repo, DeepSWE, Automation Bench, Toolathlon-Verified, EnterpriseOps-Gym, GDPval, Claw-Eval, DeepPlanning, Tau3-Bench; 'leads on most'; on JobBench it beats the baseline in most of 35 occupations (self-reported)
- Strategy: a high-capacity teacher agent first, then distillation into small models (a Nanbeige4.3-3B is named)
- Pandaily covered it on Oct 26; HF showed ~2.8k downloads and ~590 likes on Oct 26

## What happened

A recruitment-platform company's AI team released a mid-sized open model aimed at coding agents rather than chat. Weights for the
instruct and base versions went up on Hugging Face under a permissive licence, with serving examples for vLLM and SGLang.

## Why it matters

It adds a 30B-class Chinese open model with a longer context (512K) than its named rivals. Its benchmark lead over Qwen, Meta and
Google models of similar size is the team's own claim until independent leaderboards test it.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 177 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 116 days after its cutoff
- Gemini 4 Hydrogen (training cutoff March 2026): 207 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 146 days after its cutoff

## Sources

1. [Hugging Face: Nanbeige/Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) (huggingface.co, official)
2. [Pandaily: BOSS Zhipin's Nanbeige team open-sources Nanbeige4.3-33B (Oct 26)](https://pandaily.com/nanbeige4-3-33b-open-source-agent-coding-model-512k-context-apache-2-boss-zhipin) (pandaily.com, press)

## Changes

- 2026-10-26 (filed): Created (model card read on Hugging Face)

## Related

- 2026-08-10: [Meta returns to open weights with Muse Glimmer, a 30B Apache-2.0 agentic model](https://postcutoff.com/e/2026-08-10-meta-muse-glimmer-open-weights/index.md)
- Models: [Nanbeige4.3-33B](https://postcutoff.com/m/nanbeige4-3-33b/) (`nanbeige4-3-33b`)