BOSS Zhipin's Nanbeige Lab Releases a 33B Dense Model Built to Teach a Future 3B Agent

The language-model lab inside the recruitment app BOSS Zhipin has released its biggest model yet, and says the main reason it built the model was to train a much smaller one. Nanbeige4.3-33B appeared on Hugging Face over the weekend of October 24–25, 2026, together with a matching base model. Nanbeige describes itself on Hugging Face as a large language model research team at BOSS Zhipin, the jobs app run by Kanzhun. The model card calls the 33B model “the flagship teacher model for the Nanbeige4.3 series”. Its job is to anchor distillation of agentic skills into smaller models, starting with an upcoming Nanbeige4.3-3B. The team also pitches it as a “cost-effective option for immediate real-world deployment” on coding and office-style agent work.

A plain dense design with a very long window

Nanbeige4.3-33B is a dense model rather than a mixture of experts. Its configuration file describes a standard LlamaForCausalLM stack. It has 52 layers, a hidden size of 7,168, 56 attention heads sharing 8 key-value heads, and a 166,912-token vocabulary. Hugging Face counts 33.1 billion BF16 parameters across 14 safetensors shards totalling about 66 GB. All 33 billion parameters are used for every token, so it lacks the per-token compute savings of sparse designs.

The context window is the standout figure. The card says the model supports up to 524,288 tokens (512K), and the configuration sets an unusually high RoPE base frequency of 120 million to stretch positional encoding that far. The serving examples for both SGLang and vLLM use four-way tensor parallelism.

The vendor’s scoreboard

All benchmark numbers come from Nanbeige’s model card. The team compares the model with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, with every model at its highest thinking effort, using the same hyperparameters and harness for each benchmark. Its own model ran with the preserve_thinking option switched on.

The widest margins are on agentic coding. Nanbeige reports 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen3.8-27B, and 65.5 on DeepSWE against 42.5. On the harder Terminal-Bench 4.0 it reports 23.8, while the three comparison models score 6.3 or below. Gemma4-31B scores zero. On office and workflow tests the card lists 54.2 on Automation Bench, 44.8 on JobBench and 63.3 on DeepPlanning. On YC-bench, a simulated startup run, it reports $1,025,437, against $641,290 for Qwen3.8-27B.

The card doesn’t claim a clean sweep. Qwen3.8-27B comes out ahead on SWE-Bench Pro (61.7 to 60.3) and GDPval (86.4 to 84.8), and Gemma4-31B leads GPQA Diamond with 91.3 to Nanbeige’s 89.5. The team’s own summary is that the model “leads on most of the benchmarks”. Several results depend on particular harnesses, including mini-swe-agent 2.4.6, Claude Code and Terminus 2, and Writeble has not independently reproduced any of them.

Thinking that carries across turns

The release leans on the chat template. Reasoning (enable_thinking) is on by default. A second switch, preserve_thinking, keeps the model’s reasoning from earlier assistant turns in the conversation. Nanbeige recommends turning it off for ordinary chat and on for multi-turn tool use, office tasks and coding agents. Tool calls can be written as XML or JSON, and the team recommends XML “for the best tool-calling performance”.

The suggested sampling settings depend on the job. For agentic and tool-use work the card recommends temperature 1.0, top-p 0.95 and up to 65,536 new tokens. For reasoning and chat it recommends temperature 0.6 with a ceiling of 262,144 new tokens. Both server recipes use the Qwen3 reasoning parser and the qwen3_coder tool-call parser. Loading through Transformers requires trust_remote_code.

Licence listed in metadata, no LICENSE file

The model card’s metadata lists Apache-2.0. The repository contains no LICENSE file, so for now that metadata entry is the only licence statement Nanbeige has made, and the full licence text isn’t in the repo. That matters for anyone weighing commercial use. The community moved quickly anyway. By the evening of October 25 (UTC), Hugging Face listed more than a dozen third-party conversions, including GGUF, MLX and NVFP4 builds, some of them advertised as having safety behaviour removed. Those builds are not from Nanbeige.

Nanbeige hasn’t given a release date for the 3B student model the 33B was built to teach. The card asks companies interested in industry-specific agents to get in touch about joint work on data and training.