Nanbeige4.2-3B Released with Looped Transformer for Agent Capabilities
Nanbeige Lab has officially released the Nanbeige4.2 model series on the ModelScope platform. The most notable release is the Nanbeige4.2-3B, a compact Agent model based on the Looped Transformer architecture that increases model capacity without adding parameters, outperforming larger models in multiple benchmarks.
Confirmed
- Model Architecture and Parameters: Nanbeige4.2-3B has 4B total parameters and 3B non-embedding parameters. The series also includes a Looped Transformer base model trained on 28T tokens, alongside a 3B parameter SFT+RL model.
- Benchmark Performance: According to @teortaxesTex and @UsedMorning9886, the 3B model achieved a high score of 63.6 on the SWE-Bench and gained a leading edge over models like Qwen 3.5 and Gemma 4 in various Agent benchmarks.
- Future Plans: The official team has confirmed that the Nanbeige4.5 model is currently in training.
Why it matters
This model demonstrates the potential of the Looped Transformer architecture to break through the performance limits of small-parameter models. By utilizing a looping mechanism to increase capacity without scaling up parameters while achieving excellent results in Agent-related tasks, it provides a highly promising new solution for efficient, low-cost edge-side Agent deployment.
2026-07-22 ~ 2026-07-23 · 5 related posts
Primary sources
- [source] Nanbeige4.2-3B launches as a 3B Looped Transformer model that beats larger baselines — Wooden-Deer-1276 · 2026-07-22
- Nanbeige 4.2-3B posts strong benchmark scores against Qwen and Gemma — teortaxesTex · 2026-07-22
- [source] Nanbeige4.2-3B claims 63.6 on SWE-Bench Verified with a Looped Transformer — UsedMorning9886 · 2026-07-22
- Nanbeige releases 3B agent model with Looped Transformer and previews 4.5 training — udmrzn · 2026-07-22
- [source] Nanbeige4.2 Lands on ModelScope Featuring 28T-Token Looped Transformer — snwy_me · 2026-07-23