Nanbeige4.3-33B, an open 33B agent model, tops Qwen3.8-27B on most of its own tests

The Nanbeige LLM Lab has published Nanbeige4.3-33B on Hugging Face, a 33-billion-parameter open-weight model under the Apache-2.0 license. Its model card claims the best agentic scores among open models of roughly 30B size.

A 33B teacher with a 512K context window

According to the model card, Nanbeige4.3-33B handles English and Chinese text, ships in BF16 Safetensors format and accepts up to 524,288 tokens (512K) of context. It is built on a base model called Nanbeige4.3-33B-Base. The card gives no architecture details beyond tagging it as a llama and nanbeige model, and it does not describe training data, methods or compute.

The card calls it the flagship “teacher” of the Nanbeige4.3 series. The idea is to pass its agentic skills down to smaller siblings such as Nanbeige4.3-3B. Hugging Face lists 3 finetunes and 37 quantized versions, and no hosted inference provider at the time of writing.

Thinking mode is on by default and can be switched off with enable_thinking. The card recommends preserve_thinking=True for multi-turn tool use and code agents, and False for chat. For agent work it suggests a temperature of 1.0, top-p of 0.95 and up to 65,536 new tokens. The sample server commands for vLLM and SGLang use four-way tensor parallelism, which implies a multi-GPU setup. By simple arithmetic, 33 billion parameters in BF16 come to roughly 66 GB of weights, before any context cache. The card also points to quantized builds for llama.cpp, Ollama and LM Studio for smaller machines.

Benchmarks the lab ran against three rivals

The comparison covers Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all run at their highest thinking effort with the same inference settings and harnesses, the card says. These are the lab’s own numbers, and the card does not say that anyone outside the team has reproduced them.

On Terminal-Bench 2.1, run with the Terminus 2 harness, Nanbeige4.3-33B scores 80.9, against 73.0 for Qwen3.8-27B, 61.8 for Muse-Glimmer-30B and 47.2 for Gemma4-31B. The gaps widen on harder tests. On DeepSWE it posts 65.5 versus 42.5 for Qwen, and on Terminal-Bench 4.0 it scores 23.8 versus 6.3. The card also reports 54.2 on Automation Bench, where Qwen scores 38.3.

It does not win everywhere. Qwen3.8-27B leads on SWE-Bench Pro (61.7 to 60.3) and on GDPval (86.4 to 84.8), and Gemma4-31B leads on GPQA Diamond (91.3 to 89.5). On AIME 2026 Nanbeige scores 95.2. In the YC-bench business simulation the card lists $1,025,437 for Nanbeige versus $641,290 for Qwen.

The footnotes matter for anyone comparing numbers. Some tests used mini-swe-agent v2.4.6 and others used Claude Code harnesses, SWE-Atlas-QnA was graded by GPT-6 Sol against 1,304 rubrics, and GDPval used revised rubrics. Scores obtained with other harnesses or rubrics may not line up with these.

Why a big teacher for small agents is a live question

The angle here is technical: the model is explicitly a source of training signal, not only an end product. Distillation trains a smaller student on the outputs of a stronger teacher. As we explained in our distillation explainer, the technique goes back to a 2015 paper by Hinton, Vinyals and Dean, and Hugging Face’s DistilBERT in 2019 was about 40% smaller than BERT while keeping around 97% of its language understanding. Students rarely beat their teachers, so the quality of a flagship like this sets a ceiling for the small models that follow.

The lab has a track record in small models. Nanbeige LLM Lab is the model team of the Chinese recruiting platform BOSS Zhipin, whose parent is Kanzhun. According to listings of its earlier work, Nanbeige4-3B arrived in December 2025, pre-trained on 23 trillion tokens, and Nanbeige4.1-3B followed in February 2026 with a 256K window and support for long tool-calling chains of more than 500 rounds. Those details come from third-party model directories, not from this card. Against that history, a 33B release is a step up in size, with the stated aim of feeding the 3B line.

Distillation is also contested. Anthropic alleged in February 2026 that DeepSeek, Moonshot AI and MiniMax generated more than 16 million exchanges with Claude through about 24,000 fraudulent accounts. China has denied the claims. Nothing in the Nanbeige card touches on that dispute, and it says nothing about what data its teacher was trained on. That gap is typical of open-weight releases, but it means the card cannot settle where the model’s abilities come from.

Who can run it today, and what is still unproven

The Apache-2.0 license allows commercial use and modification, and the weights are free to download. Developers can load the model with Transformers, serve it with vLLM or SGLang, or try quantized builds locally. The card requires trust_remote_code=True, which means running custom code from the repository, so teams should review it first. No hosted API exists yet, so anyone without GPUs has to wait for a provider or a quantized build.

For readers choosing a model in the 30B class, the card makes Nanbeige4.3-33B a candidate for terminal and coding agents, and it is worth testing against Qwen3.8-27B on a team’s own tasks. What remains unestablished is whether independent evaluators will reproduce the margins, how the model behaves on real long-running work, and what it cost to train. The card also reports a comparison across 35 occupations on JobBench, but the details sit in a chart the lab produced itself.