Nanbeige4.3-33B is a specialized large language model (LLM) developed as a foundational flagship agentic model within the Nanbeige4.3 series. Designed for complex coding and professional cowork tasks, the model aims to serve as a high-capacity teacher to facilitate the distillation of advanced agentic capabilities into smaller, lightweight models.

Core Methodology and Technical Framework

Nanbeige4.3-33B is built on a 33-billion parameter architecture and is optimized specifically for agentic workflows, emphasizing reasoning, tool use, and multi-turn interaction.

1. Agentic Architecture: The model is trained with an explicit emphasis on "thinking effort," allowing it to generate internal reasoning steps before providing a final response. This is controlled via a configurable chat template where enable_thinking toggles the generation of explicit reasoning content, and preserve_thinking maintains reasoning context across multi-turn interactions.
2. Tool-Use Optimization: The model adopts an XML-based structured output format for tool calling ($tool_call_format="xml"$). This choice is intended to improve parseability and reliability in agentic loops, particularly for code-agent workflows and automated enterprise operations.
3. Inference and Context: The model supports a significantly large context window of up to $524,288$ tokens ($512K$), accommodating extensive historical dialogues and large-scale code repositories. The inference hyperparameters are optimized for different task categories:
   • Agentic/Tool-use tasks: Optimized with a temperature of $1.0$ and Top-p of $0.95$.
   • Reasoning/Chat tasks: Optimized with a lower temperature of $0.6$ for higher coherence.
4. Evaluative Benchmarking: Performance is measured against a suite of specialized agentic benchmarks, including Terminal-Bench, DeepSWE, SWE-Bench Pro, and JobBench. The evaluation framework emphasizes "highest thinking effort" and employs rigorous constraints, such as those found in pass@k methodologies and task-specific rubrics judged by external models (e.g., GPT-6 Sol).

Performance Highlights

Nanbeige4.3-33B exhibits state-of-the-art performance in the ~30B parameter class, significantly outperforming competitors such as Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B across core agentic metrics. Specifically:

• Agentic Coding: It shows superior capabilities in terminal automation and software engineering benchmarks (e.g., scoring 65.5 on DeepSWE and 60.3 on SWE-Bench Pro).
• Professional Cowork: The model demonstrates high versatility in task automation, achieving competitive scores in professional benchmarking environments like JobBench, where it maintains strong efficacy across 35 distinct occupations.
• Reasoning: It displays robust logical reasoning, achieving high performance on complex benchmarks such as GPQA Diamond ($89.5$) and AIME 2026 ($95.2$).

In summary, Nanbeige4.3-33B serves as a high-performance anchor model intended to provide the architectural and logic-based foundation for future model distillation and real-world deployment in industry-focused, domain-specific agentic applications.