Qwen3.5-9B-Base

Qwen3.5-9B-Base

Gives researchers and developers a pre-trained 9B base model with a vision encoder, meant for fine-tuning and experiments rather than direct chat. Handles text and images.

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: Qwen3.5-9B-Base is the pre-trained-only 9B model of the Qwen3.5 series, a language model with a vision encoder attached. It is a base model, so it is meant as a starting point for fine-tuning and research, not for direct conversation. It supports 201 languages and dialects.

Who it suits: Researchers and developers who want a strong open base to fine-tune, probe with in-context learning, or build on.

Industry | Who uses it | How they use it | Example
Research | ML researchers | Run in-context learning experiments on a base model | Test few-shot prompts without chat tuning
Software | Developers | Fine-tune it for a product task | Train a LoRA adapter for a support assistant
Localization | Language teams | Adapt it to a language or dialect | Fine-tune on text in a regional language

If you want a model that chats out of the box, you can skip it and pick an instruction-tuned release.

Where to use it:

• Hugging Face Transformers
• vLLM and SGLang
• Fine-tuning setups, including LoRA-style training
• Windows, macOS or Linux, for downloading the files

Use case | Who | What they get back
Fine-tuning | Developers | A custom model built on the 9B base
In-context experiments | Researchers | Behavior of a base model under few-shot prompts
Multimodal research | Researchers | A base that accepts text and images

How to set it up:

• Python with pip
• Hardware for about 19.3 GB in BF16, or about 4.8 GB at 4-bit by Hugging Face's estimate
• Transformers, vLLM or SGLang, all of which read these files

1. Install the Hugging Face hub package.
2. Download the model files.
3. Load them in your fine-tuning or research setup.

pip install -U huggingface_hub
hf download Qwen/Qwen3.5-9B-Base

How to use it: Load the weights in Transformers, vLLM or SGLang and use them for fine-tuning or few-shot experiments. The card says the control tokens were trained, so LoRA-style training with the official chat template does not need new embeddings.

Example: You load the base model and fine-tune a LoRA adapter on your own examples with the official chat template. The control tokens already work, so you do not have to train the embeddings.

For newcomers and experts: If you are new, know that a base model continues text and does not follow instructions well. Fine-tune it, or use an instruction-tuned sibling for chat.

The language model has 32 layers arranged as 8 groups of three Gated DeltaNet blocks followed by one Gated Attention block, with a hidden size of 4096 and multi-step MTP training. Context is 262,144 tokens natively, extensible to 1,010,000.

Pre-trained base with a vision encoder
Load in Transformers, vLLM or SGLang
Fine-tune or run in-context tests
Your adapted model

Good to know:

• It is pre-trained only, so it is not tuned for direct interaction
• The card says its intended uses are fine-tuning, in-context experiments and research or development
• License: Apache 2.0
• Cost to run: none named, you supply the hardware
• Data it reaches: your prompts, images and the files downloaded from Hugging Face
• Not for: direct chat, which the card says is not the intended use

Title: Context length in tokens
Native: 262144 tokens
Extended: 1010000 tokens

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 9.7B
• Weights: about 19.3 GB in BF16; about 4.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 287,397
• Likes: 154
• License: apache-2.0
• Quantized versions: 50
• Fine-tunes: 664
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, conversational

Links

• Model: https://huggingface.co/Qwen/Qwen3.5-9B-Base
• Files: https://huggingface.co/Qwen/Qwen3.5-9B-Base/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen3.5-9B-Base

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