openbmb/ minicpm-v4.7:moe

3,867 1 week ago

MiniCPM-V 4.7 — the latest multimodal LLM in the MiniCPM-V series, in 1B and 35B-A3B (MoE) variants. Single-image, multi-image, and video understanding, with smooth on-device MoE inference.

vision 1b 
ollama run openbmb/minicpm-v4.7:moe

Details

1 week ago

d3df93d6e613 · 22GB

model
qwen35moe
·
34.7B
·
Q4_K_M
projector
clip
·
553M
·
BF16
system
You are a helpful assistant.
params
{ "num_ctx": 8192, "stop": [ "[\"<|im_start|>\", \"<|im_end|>\"]" ], "temper
template
{{- if .Messages }}{{- range $i, $_ := .Messages }}{{- $last := eq (len (slice $.Messages $i)) 1 -}}

Readme

MiniCPM-V 4.7

GitHub  ·  Hugging Face  ·  ModelScope  ·  Cookbook  ·  Demo

MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, released in two variants — 1B (dense) and 35B-A3B (MoE, the first MoE model in the series). It builds on the family’s single-image, multi-image, and video understanding capabilities, combining stronger foundation capabilities with efficient visual encoding and on-device inference.

Highlights

  • 🔥 A New Pareto Frontier in Performance and Efficiency. The 1B variant substantially outperforms Qwen3.5 0.8B in both overall score and inference efficiency; the 35B-A3B variant achieves a higher overall score in less inference time than Qwen3.6 35B-A3B.
  • 📱 Smooth On-Device MoE Inference. Combining Lookahead Expert Scheduling (LES), fine-grained layer-wise quantization, and DSpark speculative sampling, the 35B-A3B variant runs at 13+ tokens/s on iPhone 17 Pro Max — without compromising capability.
  • 🚀 Efficient Architecture. Intra-ViT early visual compression (from LLaVA-UHD v4) with mixed 4x/16x visual token compression, to trade off detail against inference cost.
  • 💻 Broad Platform Support. iOS, Android, and HarmonyOS, with day-0 support for Rockchip, Houmo, Intel, AMD, and Arm chip platforms.
  • 🛠️ Developer Friendly. Supported by SGLang, vLLM, llama.cpp, and Ollama; fine-tuning via SWIFT and LLaMA-Factory; GGUF / BNB / AWQ / GPTQ / FP8 quantized variants available.

Available tags

Tag Quantization Size Notes
1b Q4_K_M 1.8 GB recommended 1B default
1b-q4_K_M Q4_K_M 1.8 GB
1b-q8_0 Q8_0 2.2 GB higher quality
1b-fp16 F16 3.1 GB full precision
35b-a3b Q4_K_M 22 GB recommended 35B-A3B default
35b-a3b-q4_K_M Q4_K_M 22 GB
35b-a3b-q8_0 Q8_0 38 GB higher quality
35b-a3b-fp16 F16 70 GB full precision
latest Q4_K_M 22 GB same as 35b-a3b

Quick start

”`bash ollama run openbmb/minicpm-v4.7