Nanbeige LLM Lab, the model research group inside Chinese recruiting platform BOSS Zhipin, published Nanbeige4.3-33B on Hugging Face on October 24, together with a base checkpoint, both under Apache 2.0. The lab calls it the flagship teacher model of the 4.3 series and says the plan is to distil its agentic skills into smaller students, starting with a 3B.
The specification is a plain dense transformer, not a mixture of experts. Config.json lists 52 layers, a hidden size of 7,168, 56 query heads against 8 key-value heads, an FFN intermediate size of 22,016, a 166,912-token vocabulary and a maximum of 524,288 positions. Because there is no expert routing, all 33 billion parameters run on every token; BF16 weights come to roughly 66 GB, which puts the model on multi-GPU servers rather than a single consumer card. One community 4-bit quantisation is listed at about 20 GB.
The chat template exposes three switches that matter for agents: thinking on by default, a preserve_thinking flag that carries reasoning across turns, and a tools flag that enables tool calling in an XML format. The model card's own deployment recipes use SGLang and vLLM with four-way tensor parallelism and a 262,144-token cap for reasoning work.
The benchmark table compares it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at each model's highest thinking effort, and Nanbeige wins 17 of 20 rows. The widest gaps are on long multi-step agent tasks: 80.9 against 73.0 on Terminal-Bench 2.1, 23.8 against 6.3 on Terminal-Bench 4.0, 65.5 against 42.5 on DeepSWE, 53.0 against 48.3 on NL2Repo, 61.1 against 56.4 on SkillsBench. On office-style work it reports 54.2 against 38.3 on Automation Bench, 44.8 against 34.5 on JobBench, 63.3 against 49.2 on DeepPlanning, and a simulated-company score of $1,025,437 against $641,290.
It does not win everything. Qwen3.8-27B is ahead on SWE-Bench Pro, 61.7 to 60.3, and on GDPval, 86.4 to 84.8, and a footnote says the GDPval number was scored with revised rubrics. Gemma4-31B leads GPQA Diamond at 91.3 against 89.5. Every figure in the table is vendor-reported, nobody has re-run the harness independently, and the release carries no technical report.
The strategy behind it is the more interesting part. Nanbeige has shipped small models before, most recently Nanbeige4.2-3B. This time it built the large model first and released it ahead of the small one it is meant to teach, which inverts the usual order. That only makes sense if the lab believes the scarce asset is agentic behaviour rather than parameters, and that the behaviour can be moved downward. BOSS Zhipin's core business is matching people to jobs at volume, a natural market for cheap agents that operate software instead of chatting about it.
The timing is crowded. Moonshot's Kimi K3.1 Pro is live with open weights promised by November 4, MiniMax put M3.1 Flash out under an open licence, and Tencent shipped a vision model this week. Chinese labs are publishing fast enough that a single benchmark win buys about a fortnight of attention. What Nanbeige has that most of them do not is a licence with no user cap and a size a mid-sized company can actually host.
The practical read for teams evaluating it: 33B dense at 66 GB in BF16 fits on four 24 GB cards or two 80 GB cards, and the 4-bit build will run on one workstation at reduced context. That is exactly the segment where local coding agents live, and it is where a self-reported 12-point gap over Qwen deserves the most scepticism. Reproduce DeepSWE before moving a stack onto it.
Comments (0)
Log in to join the discussion
Log InNo comments yet