Nanbeige4.3-33B
BOSS Zhipin's 33B dense agent model: 512K context, Apache 2.0 tag, and an INT4 build made with quantization-aware training.
In short. Nanbeige4.3-33B is a 33-billion-parameter text model built for coding agents and office-style tasks such as documents and spreadsheets. It comes from the language-model team at BOSS Zhipin, a Chinese recruitment platform. Its Hugging Face pages list Apache 2.0, and by our estimate a 22.6 GB 4-bit build lets it run on a single 32 GB GPU.
Summary
Nanbeige4.3-33B is the flagship of the Nanbeige4.3 series from Nanbeige LLM Lab, the language-model team at BOSS Zhipin. It appeared on Hugging Face on 24 October 2026 with an Apache 2.0 license tag. The lab calls it a teacher model: a large, capable agent that it plans to use to distill agent skills into smaller models, such as a Nanbeige4.3-3B that had not been published when we checked. It is also meant to be used on its own, for coding and for office-style "cowork" tasks.
It is a dense text model with 33.1 billion parameters, of which about 30.7 billion are not token embeddings. The configuration is a standard Llama design with 52 layers, a hidden size of 7,168, 56 attention heads sharing 8 key-value heads and a context window of 524,288 tokens, so common tools load it without custom code. The chat template has a thinking mode that is on by default and a "preserve thinking" setting that Nanbeige recommends for multi-turn tool use and coding agents. The card suggests a temperature of 1.0 for agent tasks and 0.6 for reasoning and chat, and gives vLLM and SGLang commands that use four GPUs.
In Nanbeige's own table, with the same harness for every model and the highest thinking effort, it scores 80.9 on Terminal-Bench 2.1 (Terminus 2 harness) against 73.0 for Qwen3.8-27B, 61.8 for Muse Glimmer-30B and 47.2 for Gemma 4 31B, and 65.5 on DeepSWE against 42.5 for Qwen3.8-27B. It trails Qwen3.8-27B on SWE-Bench Pro (60.3 against 61.7) and GDPval (84.8 against 86.4), and Gemma 4 on GPQA Diamond (89.5 against 91.3). These are the developer's numbers. Besides the BF16 weights (66.2 GB), the lab publishes a pretrain-only base model, FP8 (35.5 GB) and INT8 (36.0 GB) builds, and an INT4 build of 22.6 GB. For the INT4 build it used quantization-aware training, with distillation from the BF16 model and reinforcement learning on the INT4 model itself, so it is not a plain after-the-fact conversion, but the card gives no scores for it. None of the repositories contains a LICENSE file, so the Apache 2.0 terms come from the Hugging Face metadata.
Good for
- Coding agents and office-style agent tasks that need a 30B-class model with a long context
- Running on one 32 GB or larger GPU with the INT4 build, or on one 48 GB GPU with FP8, by our estimate
- Teams that want a model that loads in stock Transformers, vLLM and SGLang, since it uses the Llama architecture
- Fine-tuning and distillation work, because the base model is published too
Watch out for
- Every benchmark number is Nanbeige's own, with a different harness per test, and no scores are given for the quantized builds
- The repositories carry an Apache 2.0 tag but no LICENSE file, so confirm the terms with the developer before you ship it
- It handles text only, and the smaller models the lab plans to distill from it were not available when we checked
- In Nanbeige's own table it trails Qwen3.8-27B on SWE-Bench Pro and GDPval
- The BF16 weights are 66 GB, so a single consumer GPU needs a quantized build
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Nanbeige4.3-33B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Nanbeige4.3-33B BF16 | 33.1B | 76 GB | 38 GB | 23 GB | Gaming GPU |
| Nanbeige4.3-33B-FP8 FP8 | 33.1B | 76 GB | 38 GB | 23 GB | Gaming GPU |
| Nanbeige4.3-33B-INT8 INT8 | 33.1B | 76 GB | 38 GB | 23 GB | Gaming GPU |
| Nanbeige4.3-33B-INT4 (quantization-aware training) INT4 | 33.1B | 76 GB | 38 GB | 25 GB | Gaming GPU |
| Nanbeige4.3-33B-Base BF16 | 33.1B | 76 GB | 38 GB | 23 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Ripple-65B vs Nanbeige4.3-33B: which new agent model?
- Best open-source LLMs in 2026
- Best open-source models for coding and AI agents
- Nanbeige4.3-33B: a 33B agent model with a 23 GB 4-bit build
- Ripple-65B Preview: a dense 65B agent model under Apache 2.0
- Giant open models now ship in 4-bit: what it means for memory planning
Frequently asked questions
- What is Nanbeige4.3-33B?
- Nanbeige4.3-33B is a 33-billion-parameter text model built for coding agents and office-style tasks such as documents and spreadsheets. It comes from the language-model team at BOSS Zhipin, a Chinese recruitment platform. Its Hugging Face pages list Apache 2.0, and by our estimate a 22.6 GB 4-bit build lets it run on a single 32 GB GPU.
- Can I use Nanbeige4.3-33B commercially?
- Yes. Nanbeige4.3-33B is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Nanbeige4.3-33B open source?
- Nanbeige4.3-33B is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Nanbeige4.3-33B on my own computer?
- Yes, on a good graphics card. The smallest variant, Nanbeige4.3-33B, needs about 23 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Nanbeige4.3-33B released?
- Nanbeige4.3-33B was released by Nanbeige LLM Lab on October 24, 2026.
- How long a context window does Nanbeige4.3-33B have?
- Nanbeige4.3-33B supports a context window of up to 512K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Nanbeige4.3-33B model card (Hugging Face) huggingface.co
- Nanbeige4.3-33B-INT4 model card (Hugging Face) huggingface.co
- Nanbeige4.3-33B-Base model card (Hugging Face) huggingface.co
- Nanbeige4.3-33B-FP8 model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Muse Glimmer Meta | 29.6B | 128K | Apache 2.0 (Permissive) | 20 GB Gaming GPU | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| Ripple-65B Preview Ricursive Intelligence | 64.8B | 1M | Apache 2.0 (Permissive) | 45 GB Workstation |