TL;DR
On the evening of October 25, 2026, Nanbeige (Nanbeige LLM Lab), the large-model lab under BOSS Zhipin (Kanzhun Limited), announced via its official X account @nanbeige the open-sourcing of Nanbeige4.3-33B, a 33.1B-parameter dense language model released under Apache 2.0 with a 512K context window, claiming the strongest 30B-class agentic performance and planning to distill its skills into the upcoming Nanbeige4.3-3B.
Event Overview
On the evening of October 25, 2026 (Beijing time), Nanbeige, the large-model lab under BOSS Zhipin, announced the open-source release of Nanbeige4.3-33B via its official X account. The model is a 33.1B-parameter, BF16-precision dense language model. It was published on Hugging Face together with the base version Nanbeige4.3-33B-Base, under the Apache 2.0 license, with a 524,288-token (512K) context window, targeting agentic coding and office-style "cowork" scenarios.
Key Information
| Item | Detail |
|---|---|
| Release time | October 25, 2026 (evening, Beijing time) |
| Developer | Nanbeige LLM Lab (BOSS Zhipin) |
| Models | Nanbeige4.3-33B + Nanbeige4.3-33B-Base |
| Parameters | 33.1B (dense, non-MoE) |
| Context window | 524,288 tokens (512K) |
| License | Apache 2.0 |
| Target use | Agentic coding, office cowork |
Performance and Plans
In the lab's self-reported comparison across 20 benchmarks against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B (three ~30B-class open models), Nanbeige4.3-33B leads 17 rows: 80.9 on Terminal-Bench 2.1 (vs. 73.0 for Qwen3.8-27B), 95.2 on AIME 2026, and 54.2 on Automation Bench (vs. 38.3). The lab also stated that this model is the flagship "teacher" model of the Nanbeige4.3 series, and plans to distill agentic skills into smaller models, starting with Nanbeige4.3-3B.