Skip to main content
Back

Tsinghua's MiniCPM-V 4.6 Beats Bigger Rivals Running Fully on Your Phone

OpenBMB's MiniCPM-V 4.6 packs a 1.3B multimodal model onto your phone with 50% less visual compute and faster throughput than smaller rivals

  • MiniCPM-V 4.6 is a 1.3B vision-language model from OpenBMB/Tsinghua that runs fully on-device on iOS, Android, and HarmonyOS.
  • Scores 13 on the Artificial Analysis Intelligence Index — highest for any open-weights model under 2B parameters, beating Qwen3.5-0.8B (10) and Gemma4-E2B-it.
  • Uses intra-ViT early compression (from LLaVA-UHD v4) to cut visual encoding FLOPs by 50%+, achieving ~1.5x token throughput vs. Qwen3.5-0.8B.
  • Supports mixed 4x/16x visual token compression — switch modes at inference time for accuracy vs. speed tradeoffs.
  • Runs in 4GB GPU VRAM (full), 3GB (quantized), or 2GB CPU RAM via GGUF; works with vLLM, Ollama, llama.cpp, SWIFT, LLaMA-Factory.
  • Available free under Apache 2.0 on Hugging Face; mobile deployment code open-sourced at GitHub.

MiniCPM-V 4.6 is a 1.3-billion-parameter vision-language model built by OpenBMB , a lab jointly founded by Tsinghua University's NLP group and ModelBest , that runs fully on-device across iOS, Android, and HarmonyOS. A Tsinghua University-affiliated research team open-sourced the model, and it immediately set a new benchmark for the 1B-parameter multimodal class, outperforming Alibaba's Qwen3.5-0.8B and Google's Gemma4-E2B-it on standard vision-language tasks while delivering significantly faster inference speeds. The model is free and available under Apache 2.0.

Smaller than the competition, sharper than expected

The headline numbers are hard to ignore. MiniCPM-V 4.6 scores 13 on the Artificial Analysis Intelligence Index, sitting 3 points ahead of Qwen3.5-0.8B (score of 10) and 2 points behind Qwen3.5-2B (score of 15), establishing a new Pareto-optimal point on the Intelligence vs. Total Parameters chart. What makes that score more impressive is how it was achieved: the model used just 5.4M output tokens to run the Intelligence Index , roughly 19x fewer than Qwen3.5-0.8B's 101M, and 43x fewer than the reasoning variant's 233M.

On multimodal tasks specifically, MiniCPM-V 4.6 outperforms Qwen3.5-0.8B on most vision-language understanding benchmarks, and reaches Qwen3.5-2B-level capability on OpenCompass, RefCOCO, HallusionBench, MUIRBench, and OCRBench. Document understanding is a particular standout: the OmniDocBench gap is striking , 84.6 against Gemma4-E2B's 47.0. For anyone working with documents, PDFs, or dense text in images, that's not a marginal difference.

Benchmark comparison table across vision-language models including MiniCPM-V 4.6, Qwen3.5-0.8B, Gemma4-E2B-it, and LFM2.5-VL-1.6B

Hallucination resistance is also worth noting. HallusionBench scores 58.1 with a hallucination rate of 30.6%, compared to Qwen3.5-0.8B's 41.7% , meaning the model is significantly less likely to make things up about what it sees.

The architecture trick that makes it work

The efficiency story starts inside the vision encoder. In most multimodal models, an image gets converted into a large sequence of tokens, and the language model has to process all of them , even for simple queries. This is expensive. Global encoding produces massive token sequences, while post-ViT compression incurs the full quadratic attention cost of the vision transformer before any token reduction takes place.