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/m/nanbeige4-3-33b/

# Nanbeige4.3-33B

Nanbeige (Kanzhun / BOSS Zhipin), Nanbeige4.3 family. Status: current. Type: reasoning model. Released: 2026-10-24. Training cutoff: not published.

## How to call it

| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Hugging Face | `Nanbeige/Nanbeige4.3-33B` |  | https://huggingface.co/Nanbeige/Nanbeige4.3-33B |
| Hugging Face (base model) | `Nanbeige/Nanbeige4.3-33B-Base` |  | https://huggingface.co/Nanbeige/Nanbeige4.3-33B-Base |

## Pricing

- Price not published.
- Last checked: 2026-10-26

## Specs

- Input: text
- Output: text
- Context window: 524,288 tokens
- Max output: 262,144 tokens
- Open weights: yes
- Licence: apache-2.0

## What it has not seen

- No published training cutoff. Counted from its release (2026-10-24): at least 436 AI events since (83 major, 12 historic), as of 2026-11-13.
- Everything since September 2026: https://postcutoff.com/since/2026-09/

## What stands out

- 512K context in a ~33B open model for coding agents: Supports up to 524,288 tokens; chat template has enable_thinking and preserve_thinking (keep earlier reasoning across turns, recommended for multi-turn tool use and code agents); XML (recommended) or JSON tool calls. Source: https://huggingface.co/Nanbeige/Nanbeige4.3-33B
- Teacher model for small-agent distillation: Built as the flagship teacher of the Nanbeige4.3 series, whose agent skills are to be distilled into smaller models such as a planned Nanbeige4.3-3B. Source: https://huggingface.co/Nanbeige/Nanbeige4.3-33B

## Notes

~33.1B parameters (safetensors). Self-reported to lead Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B on most agentic coding and office benchmarks (Terminal-Bench 2.1/4.0, SWE-Atlas-QnA, NL2Repo, DeepSWE, GDPval, Tau3-Bench…); not independently verified. Recommended: temperature 1.0, top_p 0.95, up to 65,536 new tokens for agentic tasks; 0.6 and up to 262,144 for reasoning/chat. Runs on Transformers, SGLang, vLLM (Qwen3-style reasoning and tool parsers). No hosted API found.

Open-weight (Apache 2.0) agentic coding model from the Nanbeige team at Kanzhun, the company behind China's BOSS Zhipin recruitment app.

```bash
vllm serve Nanbeige/Nanbeige4.3-33B --max-model-len 524288
```

Model page: https://postcutoff.com/m/nanbeige4-3-33b/