BOSS Nanbeige Laboratory, under Direct Hire, open-sources Nanbeige4.3-33B Agent model

Article

BOSS Open-Sourcing by Zhipu 33B Agent Model: 20 items in self-reported benchmarks, 17 items, but the ultimate goal is 3B TechFlash Beating AI breaking news,BOSS Nanbeige Laboratory, a Zhipu subsidiary, has released an open-source model Nanbeige4.3-33B, focused on programming and office work Agent, supporting 512K-token context and using the Apache 2.0 License.

The official comparison includes Qwen3.8-27B、Meta Muse Glimmer 30B and Gemma 4 31B. In the published 20 tests, Nanbeige ranked first in 17 tests. Terminal-Bench 2.1 scored 80.9, higher than Qwen's 73.0; DeepSWE reached 65.5, while its competitor scored only 42.5. In the office automation test Automation Bench, it also led Qwen by 54.2, compared with Qwen's 38.3. However,SWE-Bench Pro and GDPval remain led by Qwen.

Interestingly, Nanbeige previously focused mainly on developing 3B-scale small models. Its previous-generation Nanbeige4.2-3B also improved the model's effective computational depth without increasing the parameter count by repeatedly reusing Transformer layers. This time, the team instead first trained a 33B large model, intending to have it serve as a “teacher” and then distill capabilities such as programming and tool use into the upcoming Nanbeige4.3-3B.

Original link https://m.theblockbeats.info/flash/373648

Key points

BOSS Nanbeige Laboratory, under Direct Hire, has released an open-source model Nanbeige4.3-33B。
Nanbeige4.3-33B focused on programming and office work Agent, supporting 512K context and using the Apache 2.0 license.
Among the announced 20 tests, Nanbeige ranked first in 17 tests.
Terminal-Bench 2.1, Nanbeige scored 80.9, while Qwen scored 73.0; in DeepSWE, Nanbeige scored 65.5, while the competitor scored 42.5。
In Automation Bench, Nanbeige scored 54.2, while Qwen scored 38.3。
SWE-Bench Pro and GDPval were led by Qwen.
Nanbeige Laboratory had previously focused mainly on developing 3B-scale small models; the team plans to have the 33B model serve as a “teacher,” distilling capabilities such as programming and tool calling into the forthcoming Nanbeige4.3-3B。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a route that first trains a larger 33B model and then teaches some of its capabilities to a smaller 3B model. Small models are generally better suited to reducing the resources required for deployment, but the final performance still awaits the release and validation of subsequent models.

Why it matters to readers

This indicates that the team is simultaneously focused on demonstrating large-model capabilities and pursuing a small-model deployment path; the test results reported primarily reflect the official comparisons released.

You can understand “distillation” as having a large model teach its capabilities to a small model; “Agent” refers to a model capable of performing programming or office tasks.

Risks and unknowns

The article does not provide independent institutional verification of these test results.
The article does not specify the exact release date, deployment cost, or product implementation status of Nanbeige4.3-3B.
The article provides no data on commercialization, user adoption, or financial impact.

BOSS Nanbeige Laboratory, under Direct Hire, open-sources Nanbeige4.3-33B Agent model

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center