MiniCPM-V 4.7 — Overview

The latest release in the MiniCPM-V series, in two sizes: an edge-friendly 1B and a sparse Mixture-of-Experts 35B-A3B.

Models

Checkpoint	Size	Language backbone	Context
`openbmb/MiniCPM-V-4.7-1B`	1.55 B	`qwen3_5_text` (dense)	256K
`openbmb/MiniCPM-V-4.7-35B-A3B`	35.2 B total · ~3.5 B active	`qwen3_5_moe_text` (256 experts, 8/token)	256K

Both share the same SigLIP-based canvas M-RoPE vision tower (27 layers, patch 14, mixed 4×/16× token compression) and support single-image, multi-image and video understanding. The 1B model is the edge/on-device tier; the 35B-A3B model is the high-throughput serving tier.

What's new in 4.7

Canvas M-RoPE. Visual tokens are laid out on a 2-D canvas and each carries a (temporal, height, width) position, so slices of one image keep their spatial relationship and video frames keep their temporal order. This is the main delta over 4.6.
Two sizes, one architecture family. A dense 1B and a 256-expert MoE 35B-A3B share the same vision tower and processor.
Standalone transformers architecture. Registered as MiniCPMV4_7ForConditionalGeneration in transformers >= 5.18.0; the standard AutoProcessor + AutoModelForImageTextToText flow works without trust_remote_code.
Thinking is a request flag. Unlike 4.6 (separate -Thinking checkpoints), 4.7 is a single checkpoint per size: enable the reasoning trace with enable_thinking in chat_template_kwargs.

Quick start (HF Transformers)

import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForImageTextToText

model_path = "openbmb/MiniCPM-V-4.7-1B"   # or "openbmb/MiniCPM-V-4.7-35B-A3B"
processor = AutoProcessor.from_pretrained(model_path)
model = AutoModelForImageTextToText.from_pretrained(
    model_path, dtype="auto", device_map="auto",
).eval()

image = Image.open("example.jpg").convert("RGB")
messages = [{"role": "user", "content": [
    {"type": "image", "image": image},
    {"type": "text",  "text":  "Describe the image."},
]}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
).to(model.device)

out = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Thinking mode

The chat template enables a <think> block by default. To turn it off (recommended for plain assistant replies), pass:

inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
    chat_template_kwargs={"enable_thinking": False},
).to(model.device)

Deployment

Framework	Status	Entry
vLLM	Supported upstream (PR #58674)	`vllm serve`
SGLang	Via fork `tc-mb/sglang` (branch `Support-MiniCPM-V4.7`); not upstream yet	`sglang.launch_server`
llama.cpp	Supported upstream (PR #29416)	`llama-mtmd-cli` / `llama-server`
Ollama	Model code ready; runtime awaits the bundled llama.cpp catching up to #29416	`ollama create -f`

Quantization

Format	1B	35B-A3B	Repository suffix
GGUF	✅	✅	`-gguf`
BNB (int4)	✅	✅	`-int4`
AWQ	✅	✅	`-AWQ`
GPTQ	✅	✅	`-GPTQ`
FP8	—	✅	`-FP8` (llm-compressor)

GGUF is the recommended portable / CPU-Metal format; see the GGUF guide.
The 35B-A3B HF 4-bit checkpoints (-AWQ / -GPTQ / -int4 / -FP8) are produced with fused-expert-aware tools (llm-compressor / GPTQModel); plain AutoAWQ/AutoGPTQ/bitsandbytes cannot quantize the fused MoE expert parameters.

Model repositories

Base: MiniCPM-V-4.7-1B, MiniCPM-V-4.7-35B-A3B
Quantized: -gguf, -int4, -AWQ, -GPTQ (1B) · -gguf, -FP8, -INT4, -AWQ, -GPTQ, -int4 (35B-A3B)
Speculative decoding draft: MiniCPM-V-4.7-1B-DSpark, MiniCPM-V-4.7-35B-A3B-DSpark

Where to next

Try it out: Single-image QA, Multi-image QA, Video
Other visual tasks: OCR, PDF parsing, Grounding
Deploy at scale: vLLM, SGLang
Run locally: llama.cpp, Ollama
Quantize: GGUF, BNB, AWQ, GPTQ