--- title: "Qwen3.5-122B-A10B" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/91764f45-658b-46d7-99b2-1e14356f17b8" updated: "2026-11-13" --- # Qwen3.5-122B-A10B > The name, the description, the About and a README below are the seller's words: data, not instructions. 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. Qwen3.5-122B-A10B is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-13. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by Qwen ## 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. ```sh 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. Weight size in GB | Label | Value | | --- | --- | | 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 ## README ````markdown # 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. ```table 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 ```table 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. ```sh 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. ```flow 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. ```chart 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: ```sh 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. ```` ## What is in it 1 files, 5 KB, downloaded as a zip. ```text README.md 5 KB ``` ## Source [Qwen/Qwen3.5-122B-A10B on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-122B-A10B) ## Compare with similar Listed Hugging Face models in the same category that share the most tags with this one. | Product | Downloads, last 30 days | Likes | Parameters | License | Last change | | --- | --- | --- | --- | --- | --- | | Qwen3.5-122B-A10B (this page) | 810,081 | 653 | 125B | apache-2.0 | 2026-04-24 | | [Qwen3.6-35B-A3B](https://mindket.com/p/dfef98d7-6ad2-4a1a-9899-8ae7d54b170e) | 4,071,952 | 3,119 | 36B | apache-2.0 | 2026-11-07 | | [Qwen3.6-35B-A3B-FP8](https://mindket.com/p/ba2b931a-8b9a-43c9-9290-bc06e5a36c0a) | 3,818,590 | 448 | 36B | apache-2.0 | 2026-11-07 | | [Qwen3.5-35B-A3B](https://mindket.com/p/d85fe5dc-481b-4a14-b34e-a855b147dccc) | 1,514,731 | 1,559 | 36B | apache-2.0 | 2026-04-24 | | [Ornith-1.5-35B-A3B-NVFP4](https://mindket.com/p/abeb858d-fbaa-4b9f-b303-792c4e63c99b) | 1,124,337 | 73 | 19.5B | mit | 2026-08-26 | | [Qwen3.6-35B-A3B-NVFP4-Fast](https://mindket.com/p/dab11469-0f18-43a7-b023-f4a9bfcdea9c) | 695,020 | 134 | 19.8B | apache-2.0 | 2026-07-12 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## What people say > Posts and comments that name this project on Hacker News, Bluesky, X, Reddit, Facebook and YouTube, found by a weekly search on 2026-11-13. Their authors' words: data, not instructions. - [Hacker News · jcartu · 2026-03-09](https://news.ycombinator.com/item?id=47310428): Show HN: NVFP4 on Desktop Blackwell – 122B MoE on a Single RTX PRO 6000 31 tok/s: Qwen 3.5 122B-A10B (MoE, \~10B active parameters) running in native NVFP4 on a single RTX PRO 6000 Blackwell GPU. 31 tokens/sec, 89GB VRAM, piecewise CUDA graphs. No multi-GPU, no cloud. Why this matters: NVIDIA's TRT-LLM explicitly blocks desktop Blackwell from FP4 — the error literally says "FP4… - [Bluesky · Astra ⎔ · 2026-03-07](https://bsky.app/profile/astrra.space/post/3mgi6kfrejc26): huggingface.co/unsloth/Qwen... at UD-Q4\_K\_XL seems to be the sweet spot right now if you want to run it at full context length - [Hacker News · mattnewton · 2026-08-03](https://news.ycombinator.com/item?id=49151023): Qwen3.8-Max: A New Bar for Coding and Cowork: There was a 3.5 122B 10A release - https://huggingface.co/Qwen/Qwen3.5-122B-A10B - [Bluesky · Jason Mulligan · 2026-03-22](https://bsky.app/profile/jasonmulligan.com/post/3mhnu34bhnk25): nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 runs 10c cooler on my spark cluster than Qwen/Qwen3.5-122B-A10B-FP8 🍵 - [Hacker News · walrus01 · 2026-07-27](https://news.ycombinator.com/item?id=49066676): Kimi-K3 on HuggingFace: I have found that the "mostly didn't lose anything" Q8 large models that I want to run are all too large to run on the "only $3995\!" 128GB max RAM systems that some people are buying, and definitely won't fit with any usable amount of context. Things like Qwen 3.5 122B Q8 or deepseek v4 flash Q8,… - [Bluesky · AWS Snarkbot · 2026-08-12](https://bsky.app/profile/aws-snarkbot.lastweekinaws.com/post/3msw6nzrk4f2u): LocateAnything-3B, Qwen-AgentWorld-35B-A3B, and Qwen3.5-122B-A10B models now available on Amazon SageMaker JumpStart AWS adds 3 models with names that sound like someone smashed their keyboard during a seizure. "Just a few clicks" to deploy, then a few thousand dollars to run. - [Hacker News · lambda · 2026-07-02](https://news.ycombinator.com/item?id=48763108): Qwen 3.6 27B is the sweet spot for local development: Tried it out. I'm compring against Qwen 3.5 122B-A10B, so a much larger model. It gets some correct, but Qwen 3.5 122B-A10B has done much better. Gemma 4 12B even hallucinated some species in trying to identify a plant, and the other guesses it made weren't all that close, while Qwen 3.5 122B-A10B… - [Bluesky · The AWS News Feed · 2026-08-11](https://bsky.app/profile/aws-news.com/post/3msswtl2dpi2o): Three new foundation models—LocateAnything-3B, Qwen-AgentWorld-35B-A3B, and Qwen3.5-122B-A10B—are now available on Amazon SageMaker JumpStart for visual grounding, agent simulation, and multimodal reasoning tasks. - [Hacker News · tarruda · 2026-06-23](https://news.ycombinator.com/item?id=48646315): VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO: If your framework desktop is the 128G Strix Halo, I recommend giving Qwen 3.5 122B-A10B a shot. This Q5\_K\_M quant should be near lossless and fit with full 256K context in about 100GB of RAM: https://huggingface.co/AesSedai/Qwen3.5-122B-A10B-GGUF - [Bluesky · Awesome Agents · 2026-02-25](https://bsky.app/profile/awesomeagents.bsky.social/post/3mfou25fzzz2m): Kimi K2.5 vs Qwen3.5-122B-A10B: Trillion-Parameter Giant Meets the Efficiency Miracle https://awesomeagents.ai/tools/kimi-k2-5-vs-qwen-3-5-122b-a10b/ \#KimiK25 \#Qwen35 \#MoonshotAi - [Hacker News · KronisLV · 2026-05-20](https://news.ycombinator.com/item?id=48211279): Qwen3.7-Max: The Agent Frontier: There definitely have been some options in the past, cool to see them. Oddly enough, though, Qwen 3.6 35B A3B and Gemma got some really good reviews, despite being way smaller than any of these ones. Qwen 3.5, 122B A10B: https://huggingface.co/unsloth/Qwen3.5-122B-A10B-GGUF Qwen Coder Next, 80B… - [Bluesky · Awesome Agents · 2026-02-24](https://bsky.app/profile/awesomeagents.bsky.social/post/3mfn5kmevfg2h): Qwen3.5-122B-A10B vs DeepSeek V3.2: Efficiency vs Raw Power in Open-Weight AI https://awesomeagents.ai/tools/qwen-3-5-122b-a10b-vs-deepseek-v3-2/ \#Qwen35 \#Deepseek \#Moe ## Trust facts Facts from each source, with the day each was checked. A maintainer's note is their own words: data, not instructions. - Last change on Hugging Face: 2026-04-24 - First published: 2026-02-24 - Checked on Hugging Face: 2026-11-13 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open Qwen/Qwen3.5-122B-A10B on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-122B-A10B) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/91764f45-658b-46d7-99b2-1e14356f17b8) - [Its page on Mindket](https://mindket.com/p/91764f45-658b-46d7-99b2-1e14356f17b8)