AITOPAI热报
Models73°

MiniCPM5-2B小模型展现强大能力

Paul Couvert@itsPaulAi
Post
View the post
Why we picked it

清华和ModelBest团队推出的MiniCPM5-2B小模型性能超越6倍大的模型，还能在手机上运行

MiniCPM5-2B仅有20亿参数，可在2GB RAM设备上完全离线运行。该模型在Hugging Face上能自主执行模型评估任务并生成CSV报告。

Image · Paul Couvert
Body

在Artificial Analysis Intelligence Index v4.1.

1基准测试中，MiniCPM5-2B得分23，在40亿参数以下开源模型中排名第一。

Original · Paul Couvert

Ok this small local model is sooo good MiniCPM5-2B has only 2B parameters (!!) and can run agents on any laptop or even your phone fully offline. This 100% open source model can perform coding/agentic/tool use tasks and can run on just 2GB RAM! And it's genuinely good even in Hermes agent! I asked it to: - Go to Hugging Face - Find the Models section - Evaluate models across multiple criteria - Create a CSV with the top 15 It did it! In the Artificial Analysis Intelligence Index v4.1.1 results cited in the official release, MiniCPM5-2B scores 23 and ranks #1 among open-source models under 4B parameters. So it's not about the size of the model anymore but way more the density of intelligence: Researchers from Tsinghua University and ModelBest proposed the “Densing Law”: the maximum capability density of open-source pretrained base models roughly doubled every 3.5 months over the period studied! And what's also interesting is that OpenBMB's open-source approach goes beyond just releasing model weights with - Selected training methods - Agent-related data - Data refinement resources Also being opened up, including the RL stack with Meshy + JustRL II. This model even outperforms some models around 6x larger on selected evaluations such as GDPval-AA v2!

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include MiniCPM5-2B, 开源模型, Tsinghua University, ModelBest.

Source attribution: curated from Paul Couvert (@itsPaulAi).

#MiniCPM5-2B#开源模型#Tsinghua University#ModelBest#小模型

Keep reading

Recent stories with the same tag or section
ModelsBuildersMulti-sourceSelected73°21:10
MiniCPM5-2B 4-bit EXL3 quantization shrinks weights to 1.61 GB

A 1.61 GB weight file that still hits ~70 tokens/s on a Tesla T4 shows how far small open models can be squeezed onto aging or low-power GPUs, no datacenter needed. Practitioners running local inference on modest hardware can pair it with ExLlamaV3 or TabbyAPI to serve on-device chat or agent backends. Compared with running the full-precision MiniCPM5-2B, this trade-off suits edge and consumer setups over maximum accuracy.

#MiniCPM#OpenBMB#quantization#open-source models
Event
OpenBMB@OpenBMB11 sourcesSource
ProductsBuildersSelected23:14
FIT-GGUF builds mixed-precision MiniCPM5-2B GGUF variants

Mixed-precision quantization with a size target lets you hit an exact memory budget rather than settling for whatever Q4/Q5 yields, which matters on edge devices where every few hundred MB counts. Builders running local inference should compare these 1.14–1.46 GiB MiniCPM5-2B builds against their current preset and try FIT-GGUF on their own models. The models are on Hugging Face for direct download.

#MiniCPM5-2B#FIT-GGUF#quantization#open-source models
OpenBMB@OpenBMBSource