Nanbeige4.3-33B Bets on Agents to Prepare a New Generation of Small Models

Nanbeige releases a 33-billion-parameter open model specialized in software engineering and professional agentic tasks.

To develop more compact artificial intelligence agents, the Nanbeige lab chose to start with a significantly larger model. Nanbeige4.3-33B, featuring 33 billion parameters, was designed to perform complex programming tasks, use tools, and operate in professional working environments. It will also serve as a teacher model to transfer part of its capabilities to lighter systems, including a future Nanbeige4.3-3B.

Its weights are accessible on Hugging Face under the Apache 2.0 license. The lab is publishing comparative results against Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B, showing several notable leads in benchmarks dedicated to software development agents and automation. A 33-Billion-Parameter Model Designed to Act

Nanbeige4.3-33B is a dense language model in which all parameters participate in the computation, unlike mixture-of-experts architectures that activate only a fraction of them for each token. It relies on a Llama-type architecture and features a maximum context window of 524,288 tokens, or 512K.

This capacity makes it possible to process large volumes of information within a single prompt or conversation. In particular, it can be used to inspect vast code repositories, maintain a history of a sequence of operations, or manipulate documents and instructions associated with a professional task. However, context size does not guarantee uniform comprehension of all the information it contains.

The model supports reasoning modes and tool calling. Its configuration also allows retaining reasoning traces from previous turns in long conversations, an option recommended by the lab for certain agentic workflows.

Nanbeige provides deployment instructions with Transformers, SGLang, and vLLM. These environments allow integrating the model into existing systems without relying exclusively on a proprietary inference service. High Performance in Programming Tasks

The lab compared its model to three competitors of similar size, applying the same inference parameters and execution environments for each evaluation.

On Terminal-Bench 2.1, which measures the ability of agents to solve tasks in a terminal, Nanbeige4.3-33B reaches 80.9, compared to 73.0 for Qwen3.8-27B, 61.8 for Muse-Glimmer-30B, and 47.2 for Gemma4-31B.

On DeepSWE, dedicated to agentic software development, it obtains 65.5, ahead of Qwen3.8-27B at 42.5. The gap is also pronounced on Terminal-Bench 4.0, where the published results reach 23.8 and 6.3, respectively.

However, the comparisons are not uniformly favorable to Nanbeige. On SWE-Bench Pro, Qwen3.8-27B maintains a slight lead with 61.7 against 60.3. This difference serves as a reminder that performance can vary depending on the nature of the tasks and the environment in which the models operate.

In total, Nanbeige4.3-33B leads in 17 out of the 20 evaluations presented in its model card. These figures stem from tests published by the lab and do not constitute, at this stage, an independent validation. Choices regarding agentic environments, reasoning parameters, and scoring methods also influence the results. Professional Automation Beyond Code

Nanbeige does not limit its experiments to software development. Its evaluation campaign also includes planning, tool use, automation, and professional work tasks, grouped under the term agentic cowork.

On Automation Bench, the model displays a score of 54.2, compared to 38.3 for Qwen3.8-27B. It also reaches 44.8 on JobBench, an evaluation covering tasks associated with 35 professions, compared to 34.5 for its competitor.

Results on DeepPlanning provide another indication of its abilities to handle structured operations: Nanbeige achieves 63.3 versus 49.2 for Qwen. On Toolathlon-Verified, designed to evaluate tool usage, the gap narrows to 70.4 versus 68.5.

Conversely, on GDPval, Qwen3.8-27B achieves 86.4, ahead of Nanbeige at 84.8. Performance therefore does not follow an identical hierarchy from one professional environment to another.

These benchmarks seek to measure something other than the quality of an isolated text response. An agent must sometimes read multiple resources, decide which operation to perform, call a tool, verify its output, and continue until the problem is solved. Success then depends on the model, but also on the software and control mechanisms made available to it. Building a Large Model to Train Smaller Ones

The strategy announced by Nanbeige relies on model-to-model teaching, commonly referred to as distillation. A higher-performing system is used to produce examples, solution trajectories, or training signals intended to improve the capabilities of a more compact model.

Nanbeige4.3-33B serves as the teacher model for this new family. The first lighter system mentioned is Nanbeige4.3-3B, which is expected to adopt part of the agentic capabilities developed on a larger scale.

This direction responds to concrete constraints. A model with a few billion parameters can require less memory and computing power than a 33-billion parameter model, which facilitates certain local deployments or services handling a high volume of requests.

However, Nanbeige does not yet present detailed results for Nanbeige4.3-3B. The performance of the teacher model therefore does not make it possible to determine at what level its skills can be retained after distillation. An Open Release for Agent Development

The weights of Nanbeige4.3-33B are distributed under the Apache 2.0 license, which notably permits modification and commercial uses under the conditions set out by this license.

The model supports several tool-calling formats, with a stated preference for XML in scenarios where the quality of these calls is critical. The lab also recommends distinct settings depending on whether the system is used for agentic tasks or for standard conversational exchanges.

Users thus have an executable model in their own environments, whose capabilities can be examined beyond the sole demonstrations provided by its publisher. Nanbeige also invites interested