MiniCPM5 Guide: VRAM Requirements, GGUF Builds

September 25, 2026September 27, 2026

About This Model

MiniCPM5 is OpenBMB’s series of small language models, designed from the start to run on-device—on phones and laptops. Following the 1B model, the second release is MiniCPM5-2B, a dense model with about 2.5B parameters.

The architecture is a plain Llama-style model (LlamaForCausalLM) with a 131,072-token context window. Because it needs no special implementation, common tools such as llama.cpp, Ollama and LM Studio handle it as-is. The publisher itself distributes GGUF, MLX (for Apple Silicon) and GPTQ (4-bit) builds, plus a draft model for faster inference (DSpark) and intermediate training checkpoints (Base, Midtrain, SFT).

What Makes It Stand Out

The publisher bills it as the best open model in the 2B class, with an average score of 53.9 across its comparison table. That is far ahead of the 2B-class LFM2.5-2.6B (33.2) and Qwen3.5-2B (28.0), and higher than Qwen3.5-4B (51.1), a model twice its size. Broken down by benchmark, its strengths and weaknesses are clear:

• On programming, math and tool calling, it beats Qwen3.5-4B, a model twice its size. Its 46.4 on SWE-bench Verified, which asks the model to fix bugs in real repositories, is also far ahead of the other 2B-class models (2–6).
• Long-context and agentic results are mixed. It beats the 4B model on the long-context NoLiMa, but the 4B model is ahead on AA-LCR, LongBenchPro and LongBench v2, and on the harder SWE-bench Pro (14.4 vs 28.2) and Terminal-Bench (8.6 vs 25.8).
• On the other hand, it trails the 4B model on breadth of knowledge (MMLU-Pro) and on following detailed instructions (IFEval). With fewer parameters, it can only memorize so much. It is better suited to working with code and documents you give it than to answering general-knowledge questions.

All figures are as published on the model card.

Running It Locally

• At about 2.5B parameters, it is small enough for a 4GB-class GPU or even a laptop without a discrete GPU. With the official GGUF build you can try it right away in llama.cpp, Ollama or LM Studio. Mac users can use the official MLX build.
• The 131K-token context lets you pass in long documents or whole codebases (memory use grows with context length).
• The license is Apache-2.0, which allows commercial use.

Source: openbmb/MiniCPM5-2B model card, as of 2026-09-25. Benchmark figures are as published by OpenBMB.

Our Coverage and Data

Everything Local Model Watch has published about the MiniCPM5 family: 1 article(s) covering the base model and its fine-tunes, plus converted builds we tracked after publication. Memory requirements below are computed by this site from file sizes, not quoted from model cards. Part of our model family index.

At a Glance

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 2.5B parameters

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build openbmb/MiniCPM5-2B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but openbmb/MiniCPM5-2B-GGUF provides a GGUF build you can use.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compression: the Q8_0 build measures 8.53 bits per weight — about 53% the size of the original 16-bit weights, calculated by this site from the actual file sizes.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Quantized and Converted Variants

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files.

Articles (the family’s own models first, then newest)

Repositories

• openbmb/MiniCPM5-2B

Last updated 2026-09-27 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under “Our Coverage and Data" are assembled by code from our article log.