Original title: Nanbeige/Nanbeige4.3-33B
The flagship "teacher model" of the Nanbeige4.3 series: The authors' approach is to first build a powerful agent model, then distill its agent capabilities into smaller models (such as the subsequent Nanbeige4.3-3B). It focuses on complex coding and professional office (cowork) tasks, supporting a 512K context window; the chat template allows toggling reasoning (enable_thinking), with an option to preserve previous reasoning across multiple turns (preserve_thinking), and enables tool-calling templates when tools are passed (XML format recommended, JSON also supported). Based on Nanbeige4.3-33B-Base, the repository was created on 10 24, and it was not included in previous issues.
Rising score 78.5 · Likes 1050 · Downloads 19178
Reasons
- The authors claim leading agent performance among open-source models of approximately 30B parameters: Compared against Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B using the same framework and highest reasoning setting, it achieved Terminal-Bench 2.1 80.9 (Qwen3.8-27B 73.0), DeepSWE 65.5 (42.5), Automation Bench 54.2 (38.3), and DeepPlanning 63.3 (49.2).
- Not just coding: The authors state that in most of the 35 professions covered by JobBench, it outperforms the baseline; the table shows SWE-Bench Pro 60.3 and GDPval 84.8 slightly lower than Qwen3.8-27B's 61.7 and 86.4, which the authors did not hide.
- Apache-2.0; startup commands are provided for Transformers, SGLang, and vLLM. Face-hug likes 1050 (indicating attention only).
Usage
- vLLM: vllm serve <local path> --tensor-parallel-size 4 --gpu-memory-utilization 0.8 --enable-auto-tool-choice --reasoning-parser qwen3 --tool-call-parser qwen3_coder; SGLang uses python -m sglang.launch_server --model-path <path> --tp-size 4 --reasoning-parser qwen3 --tool-call-parser qwen3_coder. Transformers requires trust_remote_code and a tokenizer with use_fast=False.
- Parameters: For agent / tool calling, use temperature 1.0, top_p 0.95, max 65,536 new tokens, and enable preserve_thinking; for reasoning and chat, use temperature 0.6, top_p 0.95, max 262,144 new tokens, and disable preserve_thinking.
Risks
- All performance metrics are self-reported by the authors, and the comparison only selected three models of the same tier; the model card does not specify the model architecture or training data, nor does it clarify whether it was trained from scratch.
- The model card does not provide VRAM requirements, and the official startup commands all use 4-card tensor parallelism; trust_remote_code is required.