minicpm-v-4.6-text-only

MiniCPM-V-4.6 Text-Only

This is a text-only extraction of openbmb/MiniCPM-V-4.6. It keeps the Qwen3.5 text backbone and tokenizer assets, and removes the vision/multimodal components.

This repository is not an official OpenBMB release.

Conversion

• Source model: openbmb/MiniCPM-V-4.6
• Source revision: d7f5ed8d15462c508d4f9d9ce5a0cf1fe8d87bcc
• Source architecture: MiniCPMV4_6ForConditionalGeneration
• Target model type: qwen3_5_text
• Target architecture: Qwen3_5ForCausalLM
• Kept tensors: 320
• Dropped tensors: 459
• Dropped components: model.vision_tower.*
• Weight format: single model.safetensors

The original VLM config stores the text backbone under text_config and the weights under model.language_model.*. The extracted checkpoint strips that prefix and writes a standalone Qwen3.5 text CausalLM config for AutoModelForCausalLM.

Validation

Validated locally with Transformers:

• AutoConfig.from_pretrained(...) loads as Qwen3_5TextConfig
• AutoTokenizer.from_pretrained(...) loads successfully
• AutoModelForCausalLM.from_pretrained(..., torch_dtype="auto", low_cpu_mem_usage=True) loads as Qwen3_5ForCausalLM
• A tiny forward pass succeeds and returns logits with shape (1, 1, 248094)
• The safetensors file contains 320 tensors
• No vision/projector/language wrapper tensor names remain
• The safetensors header matches the file size and tensor offsets

The model uses Qwen3.5 text linear-attention layers. Without optional fast-path packages, Transformers falls back to the torch implementation.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "sasa2000/minicpm-v-4.6-text-only"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [{"role": "user", "content": "Explain why the sky looks blue."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    output_ids = model.generate(**inputs, max_new_tokens=128)

print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

Limitations

This checkpoint is text-only. Image and video inputs are not supported because the vision tower, multimodal processing assets, and related preprocessing files were removed.

Please review the upstream model card and license for the original model's intended use, limitations, and terms.