What it is
Nanbeige4.3-33B is a 33B dense (non-MoE) agentic-coding model open-sourced in October 2026 by the Nanbeige Lab under BOSS Zhipin (Kanzhun). Apache 2.0, weights and a companion Base checkpoint on Hugging Face — no commercial barrier, no "contact us for commercial use" clause.
It's not a general chat model; its position is "agentic coding + office collaboration (cowork)": long context (512K, i.e. 524,288 tokens), tool calls, terminal operation. The lab calls it the 4.3 series' flagship teacher model — train a strong teacher, then distill into smaller models; next up is Nanbeige4.3-3B.
Architecture is standard LlamaForCausalLM: 52 layers, hidden dim 7,168, 56 query heads + 8 KV heads (GQA), vocab 166,912, RoPE theta 1.2e8. Being non-MoE, every token runs all 33B parameters — which decides it's unfriendly to VRAM. The chat template has three switches: enable_thinking (default on, whether to emit reasoning), preserve_thinking (keep the reasoning chain across turns — officially recommended for tool calls / coding agents), and tool calling (XML format recommended, JSON also supported).
Deployment guide
Pull weights straight from Hugging Face. The official docs give three run paths — Transformers / SGLang / vLLM:
# vLLM / SGLang both use 4-way tensor parallel (TP=4), reusing qwen3's inference parser and qwen3_coder tool-call parser
# vLLM example
vllm serve Nanbeige/Nanbeige4.3-33B --tensor-parallel-size 4 --trust-remote-code
# SGLang example
python -m sglang.launch_server --model-path Nanbeige/Nanbeige4.3-33B --tp 4VRAM math: BF16 weights are ~66 GB (33.1B × 2 bytes). That means no single card — the official recipe uses TP=4, typically 4× 24GB cards (or 2× 40GB+). For single-card, quantize: 4-bit GGUF ~20 GB runs on one 24GB card with moderate context (llama.cpp / Ollama route).
Sampling (official): - Agentic / tool calls: temperature 1.0 + top_p 0.95, up to 65,536 new tokens. - Reasoning / chat: temperature 0.6, up to 262,144 new tokens. - For multi-turn tool tasks, keep preserve_thinking on, or long-horizon reasoning breaks.
Same-tier comparison
The lab's comparison set is Qwen3.8-27B / Muse-Glimmer-30B / Gemma4-31B, all at max thinking under the same harness. Below are vendor self-scores, none independently re-tested — treat as reference:
| Benchmark | Nanbeige4.3-33B | Qwen3.8-27B | Note |
|---|---|---|---|
| Params / type | 33B dense | 27B | non-MoE vs MoE |
| Context | 512K | per peer model | — |
| Terminal-Bench 2.1 | 80.9 | 73.0 | terminal ops, biggest win |
| Terminal-Bench 4.0 | 23.8 | 6.3 | harder long-horizon, wider gap |
| DeepSWE | 65.5 | 42.5 | software agent |
| SWE-Bench Pro | 60.3 | 61.7 | loses |
| NL2Repo (Claude Code eval) | 53.0 | 48.3 | repo-level |
| Automation Bench | 54.2 | 38.3 | office automation |
| JobBench | 44.8 | 34.5 | job tasks |
| GDPval | 84.8 | 86.4 | loses (revised rubric, metrics may not match) |
| GPQA Diamond | 89.5 | 89.2 | slight win; Gemma4-31B 91.3 higher |
| AIME 2026 | 95.2 | 94.5 | reasoning close |
How to read this table: - It leads most clearly on long-horizon, multi-step agentic / coding tasks (Terminal-Bench, DeepSWE, Automation Bench are big wins) — exactly the skill its "teacher model" is meant to pass to the 3B student. - It's not a universal win: SWE-Bench Pro and GDPval lose to Qwen3.8-27B; GPQA Diamond is also slightly below Gemma4-31B. The GDPval row uses a "revised rubric" officially — discount it in cross-table comparison. - Key caveat: all numbers above are reported by Nanbeige, with no independent re-test. Early users already report weaker first-shot on 3D-generation prompts — wait for the community to run more rounds before trusting real performance. The model card also discloses no training data, method, or compute.
Sources
- Nanbeige official: Hugging Face model card Nanbeige/Nanbeige4.3-33B and -Base (open-sourced 2026-10, Apache 2.0)
- Announcement post x.com/nanbeige (2026-10-25)
- Media restatements of BF16 weights, TP=4, and the benchmark table (scores per official model card, all vendor self-test, no independent re-test)