# Nanbeige4.3-33B: A New Open Agentic Coding Model Built for Real-World Tasks
A Chinese AI research lab has introduced a powerful new language model designed to tackle both software engineering and everyday office work. The model, named Nanbeige4.3-33B, is now available for download and has quickly drawn attention for its strong benchmark performance and generous licensing terms.
## What Makes This Model Special?
Nanbeige4.3-33B is a **33 billion parameter dense language model**, meaning every parameter is active during each inference step — there is no mixture-of-experts routing. It supports an unusually long **524,288-token (512K) context window**, making it well-suited for processing lengthy documents, codebases, and multi-turn conversations.
The model is released under the **Apache 2.0 license**, which allows commercial use, modification, and redistribution with minimal restrictions. Alongside the main model, the lab also published a **base version** intended for researchers who want to apply their own custom training before deployment.
## Who Built It?
The model comes from **Nanbeige LLM Lab**, a team housed within **BOSS Zhipin**, one of China’s largest recruitment and job-matching platforms. The lab’s strategic focus is on building AI agents that can write code, automate workflows, and assist with knowledge work. Their stated roadmap involves training a strong “teacher” model first and then compressing its capabilities into smaller, more efficient student models — the first of which, a 3B parameter version, is planned but not yet released.
## Architecture Details
Under the hood, Nanbeige4.3-33B uses a configuration compatible with the widely-used `LlamaForCausalLM` architecture. Here are the key specifications:
– **52 transformer layers**
– **Hidden size of 7,168**
– **56 query heads and 8 key-value heads** using grouped-query attention (GQA)
– **Head dimension of 128**
– **Feed-forward network intermediate size of 22,016**
– **Vocabulary of 166,912 tokens**
– **RoPE base frequency of 120,000,000** for long-context handling
– **BF16 (bfloat16) weight format**
The chat template includes three important controls. A `enable_thinking` toggle activates chain-of-thought reasoning by default, which is helpful for complex tasks. The `preserve_thinking` option retains reasoning traces from earlier conversation turns — useful for multi-step tool usage. Tool calling is enabled by passing a `tools` parameter, and the lab recommends using an XML-based format for tool calls.
## Benchmark Performance
The lab evaluated the model against three other prominent open models — Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B — all run at maximum reasoning effort with shared evaluation settings.
### Agentic Coding
Nanbeige4.3-33B achieved an **80.9 score on Terminal-Bench 2.1**, significantly ahead of Qwen3.8-27B’s 73.0. On DeepSWE, the gap was even more pronounced, with scores of **65.5 versus 42.5**. On Terminal-Bench 4.0, the model scored **23.8**.
### Agentic Cowork (Office and Workflow Tasks)
On Automation Bench, it scored **54.2** compared to 38.3 for the next best competitor. JobBench results were **44.8 versus 34.5**. In a simulated earnings benchmark (YC-bench), the model achieved **$1,025,437** in simulated earnings compared to $641,290 for the closest rival.
### Reasoning
The model reported **95.2 on AIME 2026** and **84.0 on AA-LCR**, placing it near the top of the comparison group.
Out of 20 total benchmark rows, the model leads in 17. The three areas where it trails are SWE-Bench Pro (60.3 vs. Qwen’s 61.7), GDPval (84.8 vs. Qwen’s 86.4), and GPQA Diamond (89.5 vs. Gemma4-31B’s 91.3).
## Deployment Requirements
Running this model requires **multi-GPU server hardware**. The official recipes for SGLang and vLLM recommend **4-way tensor parallelism**. Based on the architecture configuration, the BF16 weights consume approximately **66 GB of GPU memory**, and the KV cache adds roughly **208 KiB per token** in BF16 precision. A full 524,288-token sequence could therefore need around **112 GB** just for KV cache storage.
One community-created quantization is currently listed on Hugging Face. As of the model’s release, no major Hugging Face Inference Providers were hosting it for direct API access.
## Recommended Settings
For agentic workflows, the lab suggests:
– **Temperature**: 1.0
– **Top-p**: 0.95
– **Max new tokens**: 65,536
For reasoning and chat interactions:
– **Temperature**: 0.6
– **Max new tokens**: 262,144
## Key Takeaways
– A 33B dense language model with 512K context, released under Apache 2.0
– Text-only input; no image or video support
– Strongest in agentic coding and workflow automation benchmarks
– Designed as the foundation for future smaller models in its series
– Base model also available for custom training pipelines
—
## Frequently Asked Questions (FAQ)
**Q: Can I use Nanbeige4.3-33B commercially?**
A: Yes. The model is released under the Apache 2.0 license, which permits commercial use, modification, and redistribution.
**Q: How many GPUs do I need to run this model?**
A: The official deployment recipes use 4 GPUs with tensor parallelism. The model weights in BF16 format require approximately 66 GB of GPU memory, plus additional memory for the KV cache depending on your context length.
**Q: Does this model support images or documents as input?**
A: No. Nanbeige4.3-33B is a text-only model. It does not process images, videos, or other multimodal inputs.
**Q: What is the difference between the main model and the base model?**
A: The main model (Nanbeige4.3-33B) is post-trained for chat and agentic tasks with thinking and tool-call capabilities. The base model (Nanbeige4.3-33B-Base) is released separately so researchers and developers can apply their own custom training before deployment.
**Q: Is there a smaller version of this model available?**
A: Not yet. The lab has announced plans to release Nanbeige4.3-3B as the first student model distilled from the 33B teacher, but no release date has been provided.
**Q: Are the benchmark scores independently verified?**
A: The scores reported here are self-reported by the Nanbeige LLM Lab using its own evaluation harnesses. Some figures differ from what competitor labs publish on their own model cards, so independent verification is still pending.
**Q: What tool-call format does the model use?**
A: The lab recommends an XML-based format for tool calls, though the specific implementation details are documented on the model card.
**Q: Can I host this model through Hugging Face Inference Endpoints?**
A: At the time of release, no Hugging Face Inference Providers were hosting the model. You would need to deploy it on your own infrastructure using the provided Transformers, SGLang, or vLLM recipes.
—
## Conclusion
Nanbeige4.3-33B represents a notable entry in the open-source agentic coding space. With its large context window, dense architecture, and Apache 2.0 licensing, it offers researchers and developers a flexible foundation for building AI agents that can handle complex coding and office tasks. Its strong showing on agentic benchmarks — particularly in terminal-based coding and workflow automation — positions it as a serious competitor among models in its size class. While some benchmark figures still need independent verification, the model’s performance and the lab’s clear roadmap toward smaller distilled versions suggest an ambitious and well-structured project.
Thank you for reading
