Qwen3.5-122B-A10B

Qwen3.5-122B-A10B

Reads images and text together and answers in text, as a mixture-of-experts model with 122B total and 10B active parameters. Suits teams with big GPU servers who want an open vision-language model for reasoning, coding and agents.

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: Qwen3.5-122B-A10B is a large open model from the Qwen team that takes images and text as input and replies in text. It has 122 billion parameters in total, with 10 billion active per token, and it supports 201 languages and dialects. The repository holds the post-trained weights in Transformers format.

Who it suits: Teams with serious GPU hardware who want an open vision-language model for reasoning, coding and agent work, under the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Software | Developers | Ask questions about screenshots, diagrams and code together | Explain an error shown in a screenshot
AI products | Agent builders | Run an agent that reads pages and images | Read a chart and report its trend
Global services | Localization teams | Serve users across 201 languages and dialects | Answer a support question in Hindi
Research | ML researchers | Study a hybrid mixture-of-experts vision-language model | Compare it with a smaller Qwen3.5 model

Skip it if you have a single consumer GPU, since the 4-bit estimate alone is about 62.5 GB.

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

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

Use case | Who | What they get back
Image questions | Developers | A text answer about what an image shows
Reasoning and coding | Engineers | Written code or step-by-step answers
Agent tasks | Agent builders | Text replies for long, tool-driven workflows
Multilingual chat | Support teams | Replies in many of the 201 languages and dialects

How to set it up: You need Python and a huge amount of memory.

• Python with pip
• Memory for the weights: roughly 250 GB at BF16 precision, or roughly 62.5 GB when squeezed to 4-bit
• A serving stack such as Transformers, vLLM, SGLang or KTransformers

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

pip install -U huggingface_hub
hf download Qwen/Qwen3.5-122B-A10B

How to use it: Load the model in one of the supported stacks, then send it a prompt that holds text and, if you like, an image. The exact calls sit in the model card, which the text given here does not reproduce, so follow the card for code.

Example: You send a photo of a whiteboard sketch and ask what the diagram describes. The model replies in text with a reading of the sketch and a short summary of its parts.

For newcomers and experts: If you are new to large models, know that this one needs a server-class machine. Start by downloading the files on a machine with enough disk, then follow the card for the first prompt.

If you know the field, the language model has 48 layers in a hybrid layout that mixes Gated DeltaNet and Gated Attention blocks, each followed by a mixture-of-experts layer. It has 256 experts, with 8 routed and 1 shared active per token, and a vision encoder. It was trained with multi-step MTP. Native context is 262,144 tokens.

Download files with hf
Load in Transformers, vLLM or SGLang
Send text and an image
Route tokens through active experts
Receive a text reply

Good to know: The card says the context is 262,144 tokens natively and can be extended up to 1,010,000. Its benchmark tables compare the model with GPT-5-mini, GPT-OSS-120B, Qwen3-235B-A22B and other Qwen3.5 sizes.

Title: Weight size in GB
BF16: 250 GB
4-bit (estimate): 62.5 GB

• License: Apache 2.0
• Cost to run: none named, apart from your own hardware
• Data it reaches: the images and text you send it and the model files on your computer, and nothing else the card names
• Not for: small machines, since the weights run to hundreds of GB

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 125B
• Weights: about 250 GB in BF16; about 62.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 810,081
• Likes: 653
• License: apache-2.0
• Quantized versions: 177
• Fine-tunes: 57
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, conversational

Links

• Model: https://huggingface.co/Qwen/Qwen3.5-122B-A10B
• Files: https://huggingface.co/Qwen/Qwen3.5-122B-A10B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen3.5-122B-A10B

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