Nanbeige4.3-33B: a free, open 33B model for coding agents, scored by its own lab

On Sunday morning, October 25, the Nanbeige lab (@nanbeige (opens in a new tab)) announced Nanbeige4.3-33B on X: "To push the performance ceiling of small agents, our strategy in Nanbeige4.3 starts with building a powerful teacher first." By mid-morning Pacific the post had about 135,000 views and 1,370 likes. The weights had quietly been on Hugging Face since the evening of October 23, Pacific time.

Who made it. Nanbeige (南北阁) is the language-model team at BOSS Zhipin, the Chinese recruiting platform run by Kanzhun. Its recent releases were 3B models (the Nanbeige4.1 and 4.2 line); 4.3 starts with this much bigger one.

What it is. A 33.1-billion-parameter model (33,118,977,024 in BF16, per Hugging Face) for agent work: coding in a terminal or repo, plus what the lab calls "cowork" tasks like office automation and planning. It's a dense model, not a mixture of experts: config.json lists a standard Llama architecture with 52 layers. It reads and writes English and Chinese, takes up to 524,288 tokens (512K) of context, and is text only; its chat template tells it that it can't process images, video or audio. The license is Apache 2.0, so commercial use is allowed. A base model, Nanbeige4.3-33B-Base, came out alongside it for anyone who wants to do their own fine-tuning.

The numbers, and whose they are. The model card compares it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all "at their highest thinking effort" with the same settings and harness. On Nanbeige's runs it scores 80.9 on Terminal-Bench 2.1 (Qwen3.8-27B: 73.0), 23.8 on the harder Terminal-Bench 4.0 (6.3) and 65.5 on DeepSWE (42.5). It isn't a clean sweep: Qwen3.8-27B edges it on SWE-Bench Pro (61.7 vs 60.3) and GDPval (86.4 vs 84.8), and Gemma4-31B wins GPQA Diamond (91.3 vs 89.5). By our count it leads all three rivals on 17 of the 20 tests listed. One coding test, SWE-Atlas-QnA, is graded by another AI model (GPT-6 Sol) rather than by passing tests.

Who it's for. Teams that want an open model they can host themselves for coding agents, with room for very long sessions. The card has quickstarts for Transformers, vLLM and SGLang, and recommends XML-style tool calls and keeping the model's earlier reasoning (preserve_thinking=True) for multi-turn agent work.

What it takes to run. The official BF16 weights are 66.2 GB, and the card's vLLM and SGLang examples split the model across four GPUs. Community quantizations appeared within hours: bartowski's 4-bit Q4_K_M file is 20.8 GB and the 8-bit Q8_0 is 35.2 GB. Those are not from Nanbeige, and a dense 33B model reads all its weights for every token, so expect it to be slower on a laptop than a mixture-of-experts model of similar size.

The honest caveat. Every score on the card is Nanbeige's own, run on its own setup, and there are no independent results yet. The lab also says this model is mainly a "teacher" for smaller models, starting with a Nanbeige4.3-3B, but gives no date. Until outside tests land, treat the benchmark table as a reason to try it on your own repo, not as proof.

What we did and didn't do: on October 25, 2026 we read @nanbeige's announcement on X, the Nanbeige4.3-33B and -Base model cards, config.json, the weight index and file list on Hugging Face (via its public API), and the file sizes of bartowski's community GGUFs. MODEL has not run the model. Not sponsored; no affiliate links.

MODEL Editors