Nabangi Releases 33B Agentic AI Model

China’s Nabangi LLM Research Institute released Nabangi 4.3-33B on Hugging Face on the 24th, presenting it as an agentic AI model for coding and office-task automation. The institute, operated by major Chinese recruitment platform BOSS Zhipin, said the open-source model has 33 billion parameters and scored higher on several coding and office-task benchmarks than similarly sized competitors, including Alibaba’s Qwen 3.8-27B. It supports English and Chinese under the Apache 2.0 license, which permits commercial use and research, as well as modifying and redistributing its weights and building products based on the model.

Nabangi 4.3-33B uses a dense architecture rather than a Mixture of Experts (MoE) design, so all parameters are active when processing input tokens. The institute said this comparatively simple execution structure makes deployment more predictable, though it may use more compute resources than MoE models. It described the model as one of the strongest open models around 30 billion parameters for agent tasks and claimed it can handle real work cost-effectively when paired with major agent execution environments.

In the institute’s own benchmark, Nabangi 4.3-33B earned the top score in 17 of 20 evaluations against Qwen 3.8-27B, Meta’s Muse-Glimmer-30B and Google’s Gemma 4-31B. For difficult evaluations, the institute said all models used the “Highest thinking effort” setting and the same inference hyperparameters. Nabangi scored 80.9 on Terminal Bench 2.1, versus 73.0 for Qwen; on DeepSWE, the scores were 65.5 and 42.5. On the office-automation benchmarks, it scored 54.2 on Automation Bench versus 38.3, and 44.8 on JobBench versus 34.5. In YC-Bench’s simulated startup, it generated $1,025,437 in revenue, compared with Qwen’s $641,290.

The model did not lead every comparison. On SWE-Bench Pro, it scored 60.3, below Qwen’s 61.7, and on GDPval it scored 84.8, below 86.4. Its AIME 2026 score was 95.2 and its AA-LCR score was 84.0. Some figures in Nabangi’s comparison table differ from competitors’ published results: it lists Gemma 4-31B at 89.5 on GPQA Diamond, while Google’s published figure is 84.3. The article says evaluation settings and execution tools can affect results, making it difficult to establish model superiority from the institute’s evaluation alone.

The model supports a context window of up to 524,288 tokens. Its model card includes instructions for running it with Transformers, SGLang and vLLM. The article says BF16-precision weights alone require about 66 GB of memory, with more needed for long-context processing, making deployment on a multi-GPU server suitable. The institute plans to use the model as a teacher for smaller-model development, building a distillation pipeline to transfer training data and agent capabilities; it is also preparing a follow-up Nabangi 4.3-3B.

China’s Nabangi LLM Research Institute released Nabangi 4.3-33B on Hugging Face on the 24th, presenting it as an agentic AI model for coding and office-task automation. The institute, operated by major Chinese recruitment platform BOSS Zhipin, said the open-source model has 33 billion parameters and scored higher on several coding and office-task benchmarks than similarly sized competitors, including Alibaba’s Qwen 3.8-27B. It supports English and Chinese under the Apache 2.0 license, which permits commercial use and research, as well as modifying and redistributing its weights and building products based on the model.

Nabangi 4.3-33B uses a dense architecture rather than a Mixture of Experts (MoE) design, so all parameters are active when processing input tokens. The institute said this comparatively simple execution structure makes deployment more predictable, though it may use more compute resources than MoE models. It described the model as one of the strongest open models around 30 billion parameters for agent tasks and claimed it can handle real work cost-effectively when paired with major agent execution environments.

In the institute’s own benchmark, Nabangi 4.3-33B earned the top score in 17 of 20 evaluations against Qwen 3.8-27B, Meta’s Muse-Glimmer-30B and Google’s Gemma 4-31B. For difficult evaluations, the institute said all models used the “Highest thinking effort” setting and the same inference hyperparameters. Nabangi scored 80.9 on Terminal Bench 2.1, versus 73.0 for Qwen; on DeepSWE, the scores were 65.5 and 42.5. On the office-automation benchmarks, it scored 54.2 on Automation Bench versus 38.3, and 44.8 on JobBench versus 34.5. In YC-Bench’s simulated startup, it generated $1,025,437 in revenue, compared with Qwen’s $641,290.

The model did not lead every comparison. On SWE-Bench Pro, it scored 60.3, below Qwen’s 61.7, and on GDPval it scored 84.8, below 86.4. Its AIME 2026 score was 95.2 and its AA-LCR score was 84.0. Some figures in Nabangi’s comparison table differ from competitors’ published results: it lists Gemma 4-31B at 89.5 on GPQA Diamond, while Google’s published figure is 84.3. The article says evaluation settings and execution tools can affect results, making it difficult to establish model superiority from the institute’s evaluation alone.

The model supports a context window of up to 524,288 tokens. Its model card includes instructions for running it with Transformers, SGLang and vLLM. The article says BF16-precision weights alone require about 66 GB of memory, with more needed for long-context processing, making deployment on a multi-GPU server suitable. The institute plans to use the model as a teacher for smaller-model development, building a distillation pipeline to transfer training data and agent capabilities; it is also preparing a follow-up Nabangi 4.3-3B.