Nanbeige4.3-33B open model tops Qwen on agent coding

Nanbeige LLM Lab, the AI research team of Chinese recruitment platform BOSS Zhipin, has released Nanbeige4.3-33B as open weights under the Apache 2.0 license. The weights went up on Hugging Face on October 24, 2026. On the lab's own tests, the 33-billion-parameter model beats Qwen3.8-27B at most agent-style coding tasks.

Open weights means anyone can download the model files and run them on their own hardware. Apache 2.0 allows commercial use with few conditions.

What Nanbeige4.3-33B is

The model card on Hugging Face calls it the "flagship teacher model" of the Nanbeige4.3 series. A teacher model is a large model used to train smaller ones, a process called distillation. Data Studios reports that Nanbeige4.3-3B is named as the first smaller "student" model. No date has been given for it.

The model is built for agentic work. That means tasks where the model plans steps and calls tools, such as a terminal or a code editor, rather than only answering a question. The card says it "excels at complex coding and professional cowork tasks."

Key facts from the card and Data Studios:

• Size: about 33 billion parameters in a dense design. Every parameter works on every token, unlike mixture-of-experts models that use only part of their weights at a time.
• Context: 524,288 tokens, the amount of text it can read at once.
• Languages: English and Chinese.
• Thinking mode: on by default. The card recommends keeping earlier reasoning in multi-turn tool use and code-agent work.

Nanbeige's earlier public models were all small, 3-billion-parameter releases, Data Studios says. Startup Fortune notes the 3B line came out earlier in 2026.

The benchmark claims

The card claims "state-of-the-art agentic performance among open models in the ~30B parameter class." It compares the model with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B.

Benchmark	Nanbeige4.3-33B	Qwen3.8-27B	Gemma4-31B
Terminal-Bench 2.1	80.9	73.0	47.2
DeepSWE	65.5	42.5	2.7
SWE-Bench Pro	60.3	61.7	51.2
NL2Repo	53.0	48.3	33.8
Automation Bench	54.2	38.3	9.7
JobBench	44.8	34.5	9.0
GPQA Diamond	89.5	89.2	91.3
AIME 2026	95.2	94.5	89.2

Startup Fortune counts a win over Qwen3.8-27B on seven of eight agentic-coding tests. Qwen stays ahead on SWE-Bench Pro, a test of fixing real software issues, and on GDPval. Gemma4-31B leads on GPQA Diamond, a graduate-level science quiz.

Every number is Nanbeige's own. Data Studios points out that the lab adapted several test setups, so the scores do not line up directly with public leaderboards. For example, the card says NL2Repo was run through Claude Code. No technical report has been published, and the training data is not disclosed.

How to run it

The card gives setup recipes for Transformers, SGLang, vLLM and Docker Model Runner. Its serving examples split the model across four GPUs. The card does not state a minimum amount of GPU memory.

Data Studios estimates the full-precision weights at about 66 GB. It also estimates that a full-length context would need about 104 GiB of extra memory for the model's working cache. Both are the outlet's own calculations, not figures from the card.

Smaller hardware is possible with quantization, which shrinks the weights by storing numbers with fewer bits. Startup Fortune reports that a 4-bit version runs on a single 24 GB graphics card. It adds that the large cache then limits how much context you can actually use. Hugging Face lists 47 community quantizations, and Data Studios found no hosted API serving the model at the time.

What this means for developers

Nanbeige4.3-33B is another strong sign that capable coding agents can run on hardware a team owns. It follows Recursive's Flint, an Apache-licensed 40B coding model released days earlier.

• Test it on your own tasks first. The scores are self-reported and use adapted harnesses. Run it on a sample of your real issues before trusting the table.
• Compare it with Qwen3.8-27B side by side. Qwen still leads on SWE-Bench Pro, which is closest to fixing issues in an existing codebase.
• Budget memory for context, not just weights. A 24 GB card can load a 4-bit copy, but long agent sessions need room for the cache.
• Follow the card's settings. Use the recommended sampling values and keep thinking enabled for tool use, as the card recommends for agent work.
• Check the license chain. Apache 2.0 covers the official weights. Community quantizations are separate uploads, so confirm each one keeps the same terms.