Nanbeige4.3-33B: a 33B agent model with a 23 GB 4-bit build
BOSS Zhipin's Nanbeige lab published Nanbeige4.3-33B on 24 October: a dense 33B text model with a 512K context, an INT4 build of 22.6 GB and an Apache 2.0 tag but no LICENSE file.
Nanbeige LLM Lab, the language-model team at the Chinese recruitment platform BOSS Zhipin, published Nanbeige4.3-33B on Hugging Face on 24 October 2026. It is a dense text model with 33.1 billion parameters, built for coding agents and office-style tasks, and its repositories carry an Apache 2.0 tag.
What is new
The lab calls it a teacher: a large agent from which it plans to distill skills into smaller models, such as a Nanbeige4.3-3B that had not been published when we checked. Its earlier small model is Nanbeige4.2-3B. The 33B model uses the standard Llama design, with 52 layers, 56 attention heads that share 8 key-value heads and a context window of 524,288 tokens, so common tools load it without custom code. Its chat template has a thinking mode that is on by default and a “preserve thinking” setting that Nanbeige recommends for multi-turn tool use.
In Nanbeige’s own table, run with the same harness for every model and the highest thinking effort, it scores 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen3.8-27B, 61.8 for Muse Glimmer 30B and 47.2 for Gemma 4 31B. 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). All of these numbers are the developer’s.
Running it
The BF16 weights are 66.2 GB. The lab also publishes FP8 (35.5 GB) and INT8 (36.0 GB) builds, a pretrain-only base model 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. By our estimate that build fits a single 32 GB GPU. The card’s vLLM and SGLang examples use four GPUs.
License
The Hugging Face metadata lists Apache 2.0, but none of the repositories contains a LICENSE file, so the terms rest on a tag. Apache 2.0 would allow commercial use with no revenue limit, but confirm the terms with the developer before you ship a product that depends on this model.
Why it matters
Nanbeige4.3-33B and Ripple-65B Preview, published eleven days apart, both report about 80 on Terminal-Bench 2.1 in their own tables, at sizes one machine can serve. Nanbeige is the smaller of the two, and by our estimate a 4-bit build of it is about half the size of a 4-bit build of Ripple. Neither table includes the other model, so the scores are not a head-to-head, and Ripple-65B vs Nanbeige4.3-33B sets out the trade-offs.
The caveats are the missing LICENSE file, text-only input, and benchmark numbers that all come from one source. For a wider list, see the best open-source models for coding and AI agents.
Sources
- 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
Models in this story
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Nanbeige4.3-33B Nanbeige LLM Lab | 33.1B | 512K | Apache 2.0 (Permissive) | 23 GB Gaming GPU | |
| Nanbeige4.2-3B Nanbeige LLM Lab | 4.17B | 256K | Apache 2.0 (Permissive) | 2.88 GB Laptop or phone | |
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Ripple-65B Preview Ricursive Intelligence | 64.8B | 1M | Apache 2.0 (Permissive) | 45 GB Workstation |