33.119B (33,118,977,024; dense)
524,288 tokens (512K)
Apache 2.0 (model-card declaration; no standalone LICENSE file in checkpoint repository)
Open-weight dense text-generation model positioned as the flagship teacher for the Nanbeige4.3 family and the foundation for distilling agentic capabilities into forthcoming smaller models such as Nanbeige4.3-3B. The BF16 checkpoint uses a Llama-compatible causal architecture with 52 transformer layers, a 7,168-dimensional hidden state, 22,016-dimensional feed-forward layers, 56 query heads, eight key/value heads, 128-dimensional attention heads, a 166,912-token vocabulary, untied input and output embeddings, and grouped-query attention. Its shipped configuration specifies 524,288 positions with a RoPE base frequency of 120,000,000. Exact parameter accounting from the published architecture is 33,118,977,024 parameters, while the 14 BF16 safetensor shards total approximately 66.24 GB. The configurable chat template enables thinking by default, can disable reasoning for an individual response, and can either discard or preserve reasoning from earlier assistant turns; Nanbeige recommends preserving prior thinking for multi-turn tool use, office tasks, and code-agent workflows while disabling it for ordinary chat and question answering. Native tool prompting supports both XML and JSON tool-call formats, with XML recommended by the publisher. The official inference guidance permits up to 65,536 generated tokens for agentic and tool-use tasks and up to 262,144 generated tokens for reasoning and chat tasks within the 512K context. Nanbeige reports that the model leads most of its shared-harness comparisons against Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B, including 80.9 on Terminal-Bench 2.1, 23.8 on Terminal-Bench 4.0, 65.5 on DeepSWE, 50.8 on SWE-Atlas-QnA, 53.0 on NL2Repo, 54.2 on Automation Bench, 44.8 on JobBench, 70.4 on Toolathlon-Verified, 63.3 on DeepPlanning, 89.5 on GPQA Diamond, 95.2 on AIME 2026, and 84.0 on AA-LCR. It reports 60.3 on SWE-Bench Pro and 84.8 on GDPval, where Qwen3.8-27B scores 61.7 and 86.4 respectively. These are publisher-run evaluations at each model's highest thinking effort rather than independent results; Nanbeige4.3-33B was evaluated with preserved historical reasoning, and individual tasks use different agent harnesses and judges including Terminus 2, mini-swe-agent, Claude Code, and GPT-6 Sol. The repository provides Transformers, SGLang, and vLLM deployment examples, but does not disclose the base checkpoint, training-token count, training corpus, supervised-finetuning recipe, reinforcement-learning procedure, or training compute.