

Nanbeige 4.2
#15 in Lütte & Edge-ModelleNanbeige · v4.2 · 2× · tolest 04. Sept. 2026
Nanbeige4.2-3B is a compact, Open Source language model from Nanbeige LLM Lab (BOSS Zhipin) with 3 billion non-embedding parameters, optimized for agentic tasks such as Tool-Nutzung, Coding, and office workflows. The model was trained from scratch on 28 trillion Token and uses a "Looped Transformer" architecture that cycles through the layer stack multiple times to increase effective model capacity without additional parameters. It supports a context window of up to 262,144 Token and significantly outperforms larger Open Source models like Qwen3.5-9B and Gemma4-12B on multiple agent and Reasoning Benchmarks according to the model card. The model is text-based (no multimodal support), freely available under Apache 2.0 license, and primarily published on Hugging Face.
Features
| Key Benchmark (%) | GPQA-Diamond: 87.4%; SWE-Bench Verified: 63.6% |
| Context Window (Tokens) | 262,144 tokens (256K) |
| License | Apache 2.0 |
| Multimodality | No, text-only (no image/audio input) |
| Platform | Hugging Face (Transformers, Safetensors); compatible with vLLM, SGLang, llama.cpp, Ollama |
| Release Date | July 2026 |