BOSS Zhipin's Nanbeige lab open-sources Nanbeige4.3-33B, a 33B dense agent model with 512K context under Apache 2.0, reporting 80.9 on Terminal-Bench 2.1 vs 73.0 for Qwen3.8-27B

Background and the problem

Open models around 30B parameters are the workhorse size for local and private agent deployments, but they have lagged well behind closed frontier models on agent work such as terminal operations, long-horizon coding and office automation. The Nanbeige LLM Lab of BOSS Zhipin (Kanzhun Limited), previously known for small models like Nanbeige4.2-3B, announced Nanbeige4.3-33B from its official account at 12:11 UTC on 10-25 (post). The strategy is to build a strong teacher model first, then distill its agent abilities into smaller models.

How it works

According to the Hugging Face model card, Nanbeige4.3-33B is a 33B dense model on a Llama-style architecture (not MoE) with up to 524,288 tokens (512K) of context, released under Apache 2.0 alongside a Base checkpoint. The chat template has two switches: enable_thinking (whether to produce reasoning) and preserve_thinking (whether to keep earlier reasoning in multi-turn conversations). The lab recommends turning on preserve_thinking for multi-turn tool use, office tasks and code agents, and recommends the XML tool-call format. It is explicitly the teacher model for the Nanbeige4.3 series, with Nanbeige4.3-3B next.

Benchmarks and measured results

All of the following come from the official model card, compared with Qwen3.8-27B using the same harness at highest thinking effort: Terminal-Bench 2.1 (Terminus 2) 80.9 vs 73.0; Terminal-Bench 4.0 23.8 vs 6.3; DeepSWE 65.5 vs 42.5; SWE-Bench Pro 60.3 vs 61.7 (behind); NL2Repo (evaluated with Claude Code) 53.0 vs 48.3; Automation Bench 54.2 vs 38.3; JobBench 44.8 vs 34.5; GDPval 84.8 vs 86.4 (behind); GPQA Diamond 89.5 vs 89.2, though below Gemma4-31B at 91.3; AIME 2026 95.2 vs 94.5. These are vendor-reported and not yet independently reproduced, and at least one early user on X reported a weak first attempt on a 3D generation prompt, so real-world performance still needs more community testing.

What developers can do now

Weights download directly from Hugging Face, and the lab documents Transformers, SGLang and vLLM setups; the SGLang and vLLM examples use 4-way tensor parallelism and reuse the qwen3 reasoning parser and qwen3_coder tool-call parser. For agent workloads the lab suggests temperature 1.0, top_p 0.95 and up to 65,536 new tokens; for reasoning and chat, temperature 0.6. Teams running coding agents locally or in private environments can compare it against Qwen3.8-27B on their own task sets before switching.