4B total / 3B non-embedding parameters
131,072 max new tokens; evaluation and assistant usage emphasize long multi-turn thinking and tool workflows
Apache-2.0
Official compact agentic release from Nanbeige, positioned as a 3B-class local personal assistant and code agent that outperforms significantly larger open models on a broad in-house and public benchmark mix. The official model card says Nanbeige4.2-3B is built on Nanbeige4.2-3B-Base and uses a Looped Transformer architecture that reuses transformer layers to increase effective capacity without adding parameters. Compared with Nanbeige4.1-3B, the release shifts emphasis from general reasoning alone toward stronger agentic behavior, with supervised fine-tuning expanded through real-world environment integrations, large-scale environment synthesis, diversified task assets and scaffolds, and trajectory- and turn-level filtering using both test-case validation and rubric-based assessment. The card also says reinforcement learning combines outcome and process rewards to stabilize compact-model training. Public benchmark tables position Nanbeige4.2-3B ahead of larger models such as Qwen3.5-9B and Gemma4-12B on multiple general-agent, code-agent, and local-assistant tasks, including GDPval rubrics, Agent-IF-Oneday, Office-QA-Pro, Pinch-Bench-V2, Claw-Gym, Claw-Eval, MCP-Atlas, SWE-Bench Verified, SWE-Bench Pro, and Terminal-Bench 2.0. The release further highlights OpenClaw-oriented local assistant use, configurable thinking and preserve_thinking controls in the chat template, XML or JSON tool-call formatting, and custom parser support across Nanbeige-maintained SGLang, vLLM, llama.cpp, and Ollama integrations. The modeling code also points forward to Nanbeige4.5, where the team says LoopSplit, depth-attentive mHC, and concatenated n-gram embeddings are already incorporated for a later 2026 release.