OpenBMB launches MiniCPM-V 4.6 with 1.3B parameters on 6GB RAM

There is a persistent belief in AI that once a model shrinks down to 1B parameters, it gets lobotomized into a toy.

ModelBest and Tsinghua University just challenged that assumption head-on.

The team alongside the OpenBMB open-source community rolled out MiniCPM-V 4.6. It is a multimodal vision-language model with just 1.3B parameters, engineered to run smoothly on edge devices with roughly 6GB of RAM.

Doing more with fewer tokens

What makes MiniCPM-V 4.6 turn heads is not just that it runs on a phone, but how efficiently it chews through visual data.

In standard evaluations, both its Instruct and Thinking variants outpaced rivals like Alibaba Qwen3.5-0.8B and Google Gemma4-E2B-it across document OCR, math reasoning, and video understanding. On the Artificial Analysis benchmark, it even edged past larger 3B models.

Here is the kicker: in Artificial Analysis testing, the non-reasoning version burned through just 5.4 million tokens. That is roughly one-nineteenth the token appetite of Qwen3.5-0.8B non-reasoning, and about 1/43rd of Qwen's reasoning build.

Basically, the model managed to get slightly bigger in parameter weight while demanding far less compute to get the job done.

How it squeezes high-res vision into edge hardware

The secret sauce lies in the LLaVA-UHD v4 architecture developed with Tsinghua. By introducing early compression for visual tokens directly inside the ViT layers, the team cut image encoding compute by 55.8%.

It also brings a hybrid visual token compression mode (4x and 16x). If you need pixel-perfect document parsing, you keep high fidelity; if you need live camera interactions on weak silicon, you dial down the token density.

When chewing on a massive 3132×3132 image, time-to-first-token clocks in at around 75.7 milliseconds—more than twice as fast as Qwen3.5-0.8B. On a single RTX 4090, throughput hits nearly 55 high-res frames per second.

The weights and code are live on Hugging Face and GitHub, backed by support for vLLM, Ollama, and llama.cpp, plus mobile test builds across iOS, Android, and HarmonyOS. A developer only needs a single consumer GPU like an RTX 4090 to fine-tune the entire stack.

Why it matters

We think this release signals a healthy pivot in edge AI from raw parameter trimming to token density.

For developers, the payoff is immediate. Running local multimodal features inside desktop apps, cars, or mobile tools used to mean steep latency penalties or crippling cloud API bills. Dropping the memory floor to 6GB while keeping full fine-tuning accessible on consumer cards makes real-world deployment practical rather than experimental.

For the broader industry, it sets a tough standard for rivals like Google and Alibaba in the sub-2B bracket. Brute-forcing smaller models with naive distillation is no longer enough; vision encoders have to be designed for compute economy from day one.

Of course, a 1.3B model will never replace frontier cloud models on deep philosophical questions or massive multi-step code synthesis. But for glanceable document scans, visual search, and low-latency interaction on everyday hardware, MiniCPM-V 4.6 proves small models can pack a serious punch.