Nanbeige releases 33B model for agentic coding and office tasks

A new flagship for agentic tasks

Nanbeige LLM Lab has launched Nanbeige4.3-33B, a dense language model specifically designed for agentic coding and office-style 'cowork' tasks. The lab, which is based within the recruitment platform BOSS Zhipin, released the model on Hugging Face under the Apache 2.0 license. This license allows for broad use and modification by the developer community.

The model is positioned as a 'teacher' model within the Nanbeige4.3 series. The lab's strategy involves building a high-performing large model first and then distilling its capabilities into smaller versions. This approach aims to create efficient models that can still perform complex reasoning and tool use.

In internal testing, Nanbeige4.3-33B outperformed several other open models in the 30B parameter range. It led 17 out of 20 benchmarks in agentic tasks, which are tasks where an AI must interact with tools or environments to complete a goal. As marktechpost.com reported, it showed particular strength in deep reasoning and automation benchmarks.

Technical architecture and context

The model utilizes a standard LlamaForCausalLM architecture with 33B dense parameters. Unlike Mixture-of-Experts (MoE) models, which only activate a portion of their parameters for each token, this model runs all 33B parameters on every token. It features 52 layers and uses grouped-query attention to manage its processing efficiently.

One of the most significant technical specifications is its 524,288-token context window. This large capacity allows the model to process and remember a vast amount of information from a single prompt, such as long technical manuals or extensive code repositories. The model also supports BF16 weights, which are a format used to balance numerical precision and memory usage during computation.

To help users interact with the model effectively, the chat template includes specific controls for reasoning. Users can toggle 'enable_thinking' to turn on reasoning steps or use 'preserve_thinking' to keep reasoning from previous turns in the conversation. The lab recommends using an XML call format when passing tools to the model for the best results.

Benchmark performance and gaps

Nanbeige compared its new model against several competitors, including Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B. In the 'agentic coding' category, it scored 80.9 on Terminal-Bench 2.1, notably higher than the 73.0 scored by Qwen3.8-27B. It also showed a significant lead in DeepSWE, scoring 65.5 compared to 42.5 for the same competitor.

The model also performed well in 'agentic cowork' tasks, which simulate office-style work. It achieved a score of 54.2 on the Automation Bench and 44.8 on JobBench. In simulated earnings tests on YC-bench, the model reached $1,025,437, significantly outperforming the $641,290 reached by its competitors in the same test.

However, the model did not lead every single category. It trailed Qwen3.8-27B on the SWE-Bench Pro benchmark, scoring 60.3 against 61.7. It also fell behind Gemma4-31B on the GPQA Diamond benchmark. These variations are often due to different harnesses, which are the specific sets of rules and tests used to evaluate the models.

Hardware and deployment needs

Running a 33B dense model requires significant hardware resources, particularly for high-speed inference. The lab provides recipes for SGLang and vLLM, which are popular frameworks for serving large language models. These recipes suggest using 4-way tensor parallelism to distribute the workload across multiple GPUs.

Based on the model's configuration, the BF16 weights require approximately 66 GB of memory. When accounting for the KV cache—the memory used to store previous parts of a conversation—the total memory requirements increase. For example, at a 524,288-token context, the KV cache becomes a significant portion of the total memory footprint.

For users looking to run the model on consumer hardware, the lab also listed a community quantization version. Quantization is a process that reduces the precision of the model's weights to make it fit on smaller GPUs. This allows the model to be more accessible to researchers and developers who do not have access to multi-GPU server clusters.

The path toward smaller models

The release of Nanbeige4.3-33B is only the first step in the lab's broader roadmap. The primary goal is to use this 33B model as a source of knowledge for smaller, more agile models. By distilling the skills of the larger model, the lab hopes to create 'student' models that are faster and cheaper to run while maintaining high performance.

The first planned student in this series is Nanbeige4.3-3B. While a specific release date for the 3B model has not been announced, the lab intends for it to inherit the agentic coding and reasoning capabilities of the 33B flagship. This distillation process is a common way to make advanced AI capabilities practical for real-world applications where low latency is required.

Developers interested in the current release can find the model and its base version on Hugging Face. These models are intended for those who need a strong 'teacher' for complex tasks or as a foundation for custom post-training. The lab continues to refine its benchmarks to ensure the models meet the demands of complex, multi-turn tool use and autonomous coding.