Nanbeige4.2 3B

Nanbeige4.2-3B is a compact instruction-tuned language model from Nanbeige Lab, released under Apache-2.0 with the card tagged for English and Chinese. Its distinguishing feature is looped depth: config.json sets num_loops=2, so the 22 transformer blocks are executed twice per forward pass — parameter count stays small while effective compute depth doubles. That is why throughput does not track what a plain 3B would suggest. The first-party repo ships 8.36 GB of bf16 safetensors; community GGUF builds put Q4_K_M at 2.57 GB and Q2_K at 1.76 GB, so the weights clear every consumer GPU this site covers, including 8 GB cards — on this model the binding constraint is context, not fit. Mainline llama.cpp supports the architecture directly: conversion/nanbeige.py registers NanbeigeForCausalLM and carries the loop parameters into the GGUF, and llama-arch.cpp / llama-model.cpp define and build LLM_ARCH_NANBEIGE. config.json declares max_position_embeddings of 262144. There is no entry in the official Ollama library; community pushes exist. Nanbeige ships no first-party GGUF — only safetensors, FP8 and GPTQ-Int8 — so quantized builds come from third-party quantizers.