Nanbeige4.3-33B

About

Nanbeige4.3-33B (Nanbeige/Nanbeige4.3-33B) is the flagship teacher model of Nanbeige LLM Lab's Nanbeige4.3 series — a 33B dense model (config.json: LlamaForCausalLM, no MoE) tuned for agentic coding and professional cowork tasks. Nanbeige's strategy is to build a high-capacity teacher agent first and anchor the distillation of advanced agentic capabilities into smaller upcoming models (e.g., Nanbeige4.3-3B).

Architecture (config.json): 52 hidden layers, hidden size 7168, dense FFN with intermediate size 22016 and SiLU activation, grouped-query-attention (GQA) with 56 attention heads and 8 KV heads at head_dim 128, vocabulary 166912, untied embeddings, bfloat16 weights. A RoPE theta of 120,000,000 (no rope scaling) supports a native context window of 524,288 tokens (512K) — among the longest in the ~30B class.

Performance: evaluated at highest thinking effort with one shared harness, Nanbeige4.3-33B leads 17 of the 20 listed benchmarks against the ~30B cohort (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) — e.g. Terminal-Bench 2.1 80.9, DeepSWE 65.5, SWE-Atlas-QnA 50.8, Automation Bench 54.2, Toolathlon-Verified 70.4, DeepPlanning 63.3 and YC-bench $1,025,437 — with strong reasoning at AIME 2026 95.2, GPQA Diamond 89.5 and AA-LCR 84.0.

Deployment: a configurable chat template exposes enable_thinking and preserve_thinking (false for chat, true for multi-turn tool use and code-agent workflows); tool calling is enabled via the tools argument with tool_call_format="xml" recommended (qwen3 reasoning/tool-call parsers in SGLang and vLLM). Recommended inference: temperature 1.0 / top_p 0.95 / 65,536 max new tokens for agentic tasks, 0.6 / 262,144 for reasoning and chat. Runs via Transformers, SGLang or vLLM (tensor parallel 4). Released October 24, 2026 under Apache 2.0, finetuned from Nanbeige/Nanbeige4.3-33B-Base. Languages: English and Chinese.