模型与实验室 4.0 · 优秀 2026-09-07 · 文章

MiniCPM5-2B:Intelligence, Performance & Price Analysis(AA 14 分 4B 开权 #1)

OpenBMB 2026-09-07 放出 MiniCPM5-2B:总参数 2.6BApache 2.0HuggingFace 公开权重131k 上下文知识截止 2025-12Artificial Analysis Intelligence Index 上拿到 14 分(v4.2/v4.3 语境),排在 4B 开权模型 #1/47,同尺寸中位数只有 6原生 reasoning 支持,定价 $0/1M tokens(自托管,只付推理算力)端侧 / 笔记本本地能跑的模型和云端大模型的差距进一步压窄;对做 on-device agent 的团队,这是当下最值得 clone 的权重之一

打开原文回到归档

MiniCPM5-2B · Intelligence, Performance & Price Analysis

Source: https://artificialanalysis.ai/models/minicpm5-2b
Reference digest: OpenClaw定时任务/ClawFeed24小时高价值一览/2026-09-08-ClawFeed24小时高价值一览.md (条目 8)
Cross-ref: 调研/2026-09-08-调研-MiniCPM5-2B-端侧开源与AA分数读法.md
HuggingFace: https://huggingface.co/openbmb/MiniCPM5-2B
GitHub: https://github.com/OpenBMB/MiniCPM
AA article: https://artificialanalysis.ai/articles/openbmb-releases-minicpm5-2b

中文概要

OpenBMB 2026-09-07 放出 MiniCPM5-2B:总参数 2.6B、Apache 2.0、HuggingFace 公开权重、131k 上下文、知识截止 2025-12。Artificial Analysis Intelligence Index 上拿到 14 分(v4.2/v4.3 语境),排在 ≤4B 开权模型 #1/47,同尺寸中位数只有 6。原生 reasoning 支持,定价 $0/1M tokens(自托管,只付推理算力)。端侧 / 笔记本本地能跑的模型和云端大模型的差距进一步压窄;对做 on-device agent 的团队,这是当下最值得 clone 的权重之一。

English Summary

OpenBMB released MiniCPM5-2B on 2026-09-07: 2.6B total params, Apache 2.0, HuggingFace open weights, 131k context, knowledge cutoff 2025-12. Artificial Analysis Intelligence Index score 14 (under 4B open-weight #1 of 47; median 6). Native reasoning support, $0/M-token self-hosted pricing. The on-device/small-model ceiling is being pushed monthly; for on-device agent teams this is the most worth-cloning weights right now.

一句话

MiniCPM5-2B(2.6B、Apache-2.0)AA Intelligence Index 14 分,≤4B 开权 #1/47,中位数 6,自托管 $0/1M tokens