MiniCPM

MiniCPM is the small-model family from OpenBMB (面壁智能), a Chinese lab that has built a reputation for squeezing surprising capability out of very compact models. In our catalog the family currently has a single member, MiniCPM5-1B, released in 2026 — and it holds a special distinction: it is the smallest model we track anywhere on this site.

That positioning matters. Most of the VRAM anxiety around local LLMs comes from 7B–70B models that need careful GPU shopping. MiniCPM goes the opposite direction: a 1.08B dense model that runs on essentially anything, including CPU-only machines, while still offering a 128K context window and a permissive Apache-2.0 license. If you want a local model for a laptop, a Raspberry-Pi-class box, an old desktop, or an always-on home server, this family is where the conversation starts.

Size ladder

• MiniCPM5-1B (1.08B dense, 128K context) — Q4 weights are only about 0.6 GB, and a full Q4 @ 8K setup lands around 1.8 GB of VRAM → comfortably inside even the 12GB tier, and honestly it doesn't need a GPU at all. FP16 weights are roughly 2.2 GB. See the model page.

There is no bigger sibling in the catalog yet. If OpenBMB ships larger MiniCPM5 variants, we'll add them here — the family's history suggests more sizes are plausible over time.

What makes the family interesting

It runs everywhere. At 1.8 GB for Q4 @ 8K, MiniCPM5-1B fits in the memory of devices that haven't been called "GPU-capable" in a decade. Integrated graphics, old gaming cards, office desktops, CPU-only servers — all fine. This makes it a realistic choice for offline tools, embedded helpers, and privacy-sensitive utilities that should just work on whatever hardware is nearby.

128K context at 1B scale. A 128K context window is unusual at this size. Be aware that long context is not free: KV cache grows with context length, so pushing toward the full 128K raises the Q4 memory requirement to roughly 5 GB. Still trivially affordable on any 12GB card, but worth knowing if you plan to run it on very constrained hardware with long documents.

Dense, not MoE. The full 1.08B parameters are active on every token. There is no expert routing to reason about — what you see is what you get, and speed scales predictably with your hardware.

Apache-2.0 license. Fully permissive open licensing: fine for commercial use, fine to modify, fine to embed in products. No registration gates or usage restrictions to worry about.

Practical expectations. A model this small is not a coding assistant or a deep reasoner. Where it shines is lightweight, high-volume work: classification, summarization of short texts, autocomplete-style drafting, on-device chat for simple queries, and as a draft model for speculative decoding alongside a larger LLM. Treat it as a utility, not a brain.

Which one should you pick

There's only one model, so the real question is which hardware tier makes sense:

• No GPU / CPU-only box — genuinely viable. Q4 weights are 0.6 GB; any modern CPU with a few GB of free RAM will run it. Expect modest tok/s, but for batch jobs or low-traffic tools that's often acceptable.
• 12GB tier (RTX 3060 12GB, used RTX 2080 Ti, etc.) — total overkill in the good way. You'll have room for the full 128K context, FP16 if you want maximum quality, and several other models resident at the same time.
• 16GB and above — no benefit specific to this model, but if MiniCPM5-1B is a sidekick to a larger main model, the extra VRAM goes to the big one. See our best GPU picks for cards that make sense as a primary local-LLM GPU.

The honest summary: if you're shopping for a GPU because of MiniCPM, don't. Buy hardware for the biggest model you plan to run, and MiniCPM5-1B will happily live in whatever space is left over.

Links

• MiniCPM5-1B model page — full VRAM math at 4K/8K/32K and estimated tok/s by GPU
• MiniCPM5-1B on Hugging Face
• Best GPUs for local LLMs — our current picks by budget
• Guides — VRAM math, quantization, and large-model strategies