Nanbeige 4.1 3BNanbeige 4.1 3B

Overview

Nanbeige 4.1 3B is a 3 B dense decoder from the Nanbeige team, released February 2026. At this size it competes with Llama 3.2 3B, SmolLM3 3B, and Qwen3-4B in the edge-deployment tier. The distinctive feature per the config.json is the 262 K native context window, which is dramatically larger than any other 3B-class model in this gallery (Llama 3.2 3B ships 128 K, Qwen3-4B ships 128 K).

Architecture at a Glance

256K Context at 3B

Serving a 256 K context window on a 3 B model is unusual because the KV cache cost scales linearly with context length while the weight cost stays fixed. At 64 KiB/token × 256 K ≈ 16 GiB of KV cache, which exceeds the ≈ 6 GB bf16 weight footprint by almost 3×. The Nanbeige team's bet is that edge deployments with abundant GPU RAM but tight per-token inference cost (e.g. long-document summarization on a workstation GPU) benefit from this tradeoff.

Verdict: The Long-Context Edge Model

Nanbeige 4.1 3B is the longest-context 3B-class open weight model in the gallery. Architecturally it is conservative (plain GQA, no SWA, no thinking mode). The value is the context length: if your workload is 'summarize very long documents cheaply on a single GPU', this is the default pick. For general-purpose 3B use, Llama 3.2 3B and Qwen3-4B are stronger and have broader tooling support.