Qwen3.5-4B-Base

Qwen3.5-4B-Base

Gives researchers a pre-trained-only vision-language base model for fine-tuning and in-context learning experiments, not for direct chat. Suits developers who want a small Qwen3.5 starting point to adapt with LoRA.

Made by Qwen on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3.5-4B-Base is the pre-trained-only base model of a small Qwen3.5 model from the Qwen team. It takes images and text, has a vision encoder and a causal language model, and supports 201 languages and dialects. The card says it is meant for fine-tuning and research, not for talking to it directly.

Who it suits: Developers and researchers who want a small open base to fine-tune or test, under the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Research | ML researchers | Run in-context learning experiments on a small base model | Test how it handles a few examples in the prompt
Software | Developers | Fine-tune it with LoRA for a narrow task | Adapt it to read product photos
AI products | Model builders | Start a custom assistant from a base model | Train a support helper on private data
Education | Students | Study how a vision-language base model behaves | Compare outputs before and after fine-tuning

Skip it if you want a ready chat assistant, since this repository holds a base model without the instruction tuning that chat needs.

Where to use it: The card says the files work with several stacks.

• Hugging Face Transformers
• vLLM and SGLang
• Windows, macOS or Linux, with Python for the download

Use case | Who | What they get back
LoRA fine-tuning | Developers | A small adapter trained on your data
In-context experiments | Researchers | Completions that show how the base model follows examples
Research and development | Teams | A starting point for new models

How to set it up: You need Python and a few GB of memory.

• Python with pip
• Memory for the weights: roughly 9.3 GB at BF16 precision, or roughly 2.3 GB when squeezed to 4-bit
• A stack such as Transformers, vLLM or SGLang

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Load them in Transformers, vLLM or SGLang.

pip install -U huggingface_hub
hf download Qwen/Qwen3.5-4B-Base

How to use it: Load the model in a supported stack and fine-tune it, or feed it prompts for research. The card notes that the control tokens, im_start and im_end, were trained, so LoRA-style tuning with the official chat template needs no embedding changes.

Example: You prepare a small set of labelled examples for one task and fine-tune a LoRA adapter on this base model with the official chat template. The adapter then sits on top of the base weights and handles that task.

For newcomers and experts: If you are new to base models, know that this one continues text rather than answering questions politely. Download it, try a short prompt, and plan to fine-tune before real use.

If you know the field, the language model has 32 layers in a hybrid layout of Gated DeltaNet and Gated Attention blocks with feed-forward layers. The hidden size is 2560, and the output embeddings are tied to the token embeddings. It was trained with multi-step MTP. Native context is 262,144 tokens.

Download files with hf
Load in Transformers, vLLM or SGLang
Fine-tune with LoRA or test prompts
Use the adapted model

Good to know: The card says context is 262,144 tokens natively and can be extended up to 1,010,000. Hugging Face lists the size as 4.7B parameters, while the card gives the language model as 4B.

Title: Weight size in GB
BF16: 9.3 GB
4-bit (estimate): 2.3 GB

• License: Apache 2.0
• Cost to run: none named, apart from your own hardware
• Data it reaches: the model files on your computer and any data you fine-tune it with
• Not for: direct interaction, since the card says this base model is meant for fine-tuning and research

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 4.7B
• Weights: about 9.3 GB in BF16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 556,717
• Likes: 129
• License: apache-2.0
• Quantized versions: 38
• Fine-tunes: 276
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, conversational

Links

• Model: https://huggingface.co/Qwen/Qwen3.5-4B-Base
• Files: https://huggingface.co/Qwen/Qwen3.5-4B-Base/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen3.5-4B-Base

Then follow the usage in its model card. Nothing in this kit runs by itself.