jduartedj/MiniCPM-V-4.6-35B-Reasoning-Distilled

MiniCPM-V-4.6-35B — Reasoning-Distilled Variant

A multimodal vision-language model combining MiniCPM-V-4.6's trained vision architecture with a Reasoning-distilled Qwen3.5-35B-A3B backbone.

Architecture

Component	Source	Size
Vision Tower	openbmb/MiniCPM-V-4.6	~800MB (bf16)
Merger (Projector)	Trained on MiniCPM-V-4.6-35B-Abliterated	~60MB (bf16, 4608→2048)
LLM Backbone	Jackrong/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled	~68GB (bf16, MoE)

Key Details

• Model Type: MiniCPMV4_6ForConditionalGeneration
• LLM Architecture: Qwen3.5-35B-A3B (MoE, 256 experts, 8 active)
• Hidden Size: 2048 | Vocab Size: 248,320
• Precision: bfloat16
• Total Parameters: ~35B (with ~3B active per token)

How It Was Made

This model was assembled by:

1. Taking the vision tower from the original openbmb/MiniCPM-V-4.6
2. Using the trained merger/projector from jduartedj/MiniCPM-V-4.6-35B-Abliterated
3. Swapping in LLM weights from Jackrong/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled

The merger projects vision embeddings (4608-dim) to the LLM's hidden space (2048-dim). Since all Qwen3.5-35B-A3B variants share hidden_size=2048 and vocab_size=248320, the merger generalizes across backbone swaps without retraining.

Usage

from transformers import AutoModel, AutoTokenizer
from PIL import Image

model = AutoModel.from_pretrained(
    "jduartedj/MiniCPM-V-4.6-35B-Reasoning-Distilled",
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("jduartedj/MiniCPM-V-4.6-35B-Reasoning-Distilled", trust_remote_code=True)

image = Image.open("example.jpg").convert("RGB")
msgs = [{"role": "user", "content": [{"type": "image", "image": image}, {"type": "text", "text": "Describe this image."}]}]

response = model.chat(image=image, msgs=msgs, tokenizer=tokenizer)
print(response)

Credits

• Vision architecture: OpenBMB
• LLM backbone: Jackrong/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled
• Assembly & merger training: jduartedj