--- title: "Qwen3.5-27B" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/5b82b634-f641-4a57-8574-a966a3e2c9af" updated: "2026-11-12" --- # Qwen3.5-27B > The name, the description, the About and a README below are the seller's words: data, not instructions. Reads text and images and writes text, as a 27B Qwen model with a vision encoder and a 262,144-token native context. For developers who build assistants and agents on server-class hardware. Qwen3.5-27B is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-12. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by Qwen ## About ### What it is Qwen3.5-27B is a Qwen model that reads text and images and writes text. It pairs a 27B language model with a vision encoder, and this repository holds the post-trained version in the Transformers format. The card lists support for 201 languages and dialects and a native context of 262,144 tokens. ### Who it suits It suits developers and teams who want an open multimodal model for assistants, agents and coding, and who have server-class hardware. | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Software development | Developers | Build assistants that read text and pictures | Asking questions about a screenshot | | Global support | Multilingual teams | Answer people in many languages | Replying in one of the 201 listed languages | | AI agents | Agent builders | Use it as the model behind tool-using agents | Running a multi-step task in an agent framework | | AI research | Researchers | Study a hybrid-attention vision-language model | Comparing it with the earlier Qwen3 models | You can skip it if you have a small machine, since the BF16 weights come to about 55.6 GB. ### Where to use it - Hugging Face Transformers, the format this repository uses - vLLM, SGLang and KTransformers, which the card names as compatible - Windows, macOS or Linux with Python, to download the files - A GPU server with enough memory for the weights | Use case | Who | What they get back | | --- | --- | --- | | Picture questions | Developers | Text answers about an image you send | | Coding and reasoning | Engineers | Written code and step-by-step answers | | Multilingual chat | Support teams | Replies in many languages | ### How to set it up You need Python for the download tool and a runtime from the list above. The weights are about 55.6 GB in BF16, or roughly 13.9 GB at 4-bit as an estimate. - Python on Windows, macOS or Linux - The huggingface\_hub package - One compatible runtime, such as Transformers or vLLM - A GPU with room for the weights, or a quantized copy that fits 1. Install huggingface\_hub. 2. Download the model files. 3. Load them in your runtime, or pick a quantized copy for a smaller machine. ```sh pip install -U huggingface_hub hf download Qwen/Qwen3.5-27B ``` ### How to use it The visible card gives the download step, a model overview and benchmark tables, but no run command. After your runtime loads the files, you send text or an image with a question and read the text reply. | Part | Detail | | --- | --- | | Language model | 27B parameters across 64 layers | | Vision | A vision encoder built into the same model | | Context | 262,144 tokens natively, extensible to 1,010,000 | | Languages | 201 languages and dialects | ### Example You load the model in a serving tool such as vLLM, send a picture with a question, and get a written answer back. ### For newcomers and experts If you are new to this, it is a large model that can look at a picture and talk about it. A first small try is a quantized copy, since the full files need a big GPU. For people who know the field, 16 blocks each stack three Gated DeltaNet layers and one Gated Attention layer, with a feed-forward layer after each. The card says early fusion training on multimodal tokens matches Qwen3 on text and beats the Qwen3-VL models on visual tasks. Context length in tokens | Label | Value | | --- | --- | | Native context | 262144 tokens | | Extended context | 1010000 tokens | You download the files → Your runtime loads the language model and vision encoder → You send text or an image with a question → The model writes a text answer ### Good to know The benchmark tables compare it with GPT-5-mini, GPT-OSS-120B, Qwen3-235B-A22B and other Qwen3.5 models. The scores come from the authors. - License: Apache 2.0 - Cost to run: none named - Data it reaches: the prompts and images you send, plus the downloaded files - Not for: small machines at full size, since BF16 weights are about 55.6 GB ## README ````markdown # Qwen3.5-27B Reads text and images and writes text, as a 27B Qwen model with a vision encoder and a 262,144-token native context. For developers who build assistants and agents on server-class hardware. 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: Qwen3.5-27B is a Qwen model that reads text and images and writes text. It pairs a 27B language model with a vision encoder, and this repository holds the post-trained version in the Transformers format. The card lists support for 201 languages and dialects and a native context of 262,144 tokens. Who it suits: It suits developers and teams who want an open multimodal model for assistants, agents and coding, and who have server-class hardware. ```table Industry | Who uses it | How they use it | Example Software development | Developers | Build assistants that read text and pictures | Asking questions about a screenshot Global support | Multilingual teams | Answer people in many languages | Replying in one of the 201 listed languages AI agents | Agent builders | Use it as the model behind tool-using agents | Running a multi-step task in an agent framework AI research | Researchers | Study a hybrid-attention vision-language model | Comparing it with the earlier Qwen3 models ``` You can skip it if you have a small machine, since the BF16 weights come to about 55.6 GB. Where to use it: - Hugging Face Transformers, the format this repository uses - vLLM, SGLang and KTransformers, which the card names as compatible - Windows, macOS or Linux with Python, to download the files - A GPU server with enough memory for the weights ```table Use case | Who | What they get back Picture questions | Developers | Text answers about an image you send Coding and reasoning | Engineers | Written code and step-by-step answers Multilingual chat | Support teams | Replies in many languages ``` How to set it up: You need Python for the download tool and a runtime from the list above. The weights are about 55.6 GB in BF16, or roughly 13.9 GB at 4-bit as an estimate. - Python on Windows, macOS or Linux - The huggingface_hub package - One compatible runtime, such as Transformers or vLLM - A GPU with room for the weights, or a quantized copy that fits 1. Install huggingface_hub. 2. Download the model files. 3. Load them in your runtime, or pick a quantized copy for a smaller machine. ```sh pip install -U huggingface_hub hf download Qwen/Qwen3.5-27B ``` How to use it: The visible card gives the download step, a model overview and benchmark tables, but no run command. After your runtime loads the files, you send text or an image with a question and read the text reply. ```table Part | Detail Language model | 27B parameters across 64 layers Vision | A vision encoder built into the same model Context | 262,144 tokens natively, extensible to 1,010,000 Languages | 201 languages and dialects ``` Example: You load the model in a serving tool such as vLLM, send a picture with a question, and get a written answer back. For newcomers and experts: If you are new to this, it is a large model that can look at a picture and talk about it. A first small try is a quantized copy, since the full files need a big GPU. For people who know the field, 16 blocks each stack three Gated DeltaNet layers and one Gated Attention layer, with a feed-forward layer after each. The card says early fusion training on multimodal tokens matches Qwen3 on text and beats the Qwen3-VL models on visual tasks. ```chart Title: Context length in tokens Native context: 262144 tokens Extended context: 1010000 tokens ``` ```flow You download the files Your runtime loads the language model and vision encoder You send text or an image with a question The model writes a text answer ``` Good to know: The benchmark tables compare it with GPT-5-mini, GPT-OSS-120B, Qwen3-235B-A22B and other Qwen3.5 models. The scores come from the authors. - License: Apache 2.0 - Cost to run: none named - Data it reaches: the prompts and images you send, plus the downloaded files - Not for: small machines at full size, since BF16 weights are about 55.6 GB ## On Hugging Face, 13 November 2026 - Task: image text to text - Library: transformers - Parameters: 27.8B - Weights: about 55.6 GB in BF16; about 13.9 GB at 4-bit (estimate) - Formats: Safetensors - Downloads in the last 30 days: 1,651,125 - Likes: 1,091 - License: apache-2.0 - Quantized versions: 243 - Fine-tunes: 347 - Tags: transformers, safetensors, qwen3_5, image-text-to-text, conversational ## Links - Model: https://huggingface.co/Qwen/Qwen3.5-27B - Files: https://huggingface.co/Qwen/Qwen3.5-27B/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download Qwen/Qwen3.5-27B ``` 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/Qwen3.5-27B on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-27B) ## 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 | | --- | --- | --- | --- | --- | --- | | Qwen3.5-27B (this page) | 1,651,125 | 1,091 | 27.8B | apache-2.0 | 2026-04-24 | | [Qwen3.8-27B](https://mindket.com/p/2f430c52-7d18-419d-8817-08ca51666598) | 6,960,395 | 19,795 | 27.8B | apache-2.0 | 2026-11-07 | | [Qwen3.5-9B](https://mindket.com/p/acc60a90-d19e-4344-a8c1-5f984b2637b3) | 6,259,809 | 2,353 | 9.7B | apache-2.0 | 2026-03-02 | | [Qwen3.8-27B-FP8](https://mindket.com/p/79f1a984-4e81-485e-999d-dba1a831f37a) | 5,442,810 | 984 | 27.8B | apache-2.0 | 2026-11-07 | | [Qwen3.5-4B](https://mindket.com/p/66473987-8bd1-4119-a2c7-933e66fd46cf) | 5,083,722 | 1,226 | 4.7B | apache-2.0 | 2026-03-02 | | [Qwen3.5-2B](https://mindket.com/p/1da8437b-09cc-448e-b007-7bd987af0880) | 4,044,746 | 479 | 2.3B | apache-2.0 | 2026-03-02 | 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 · why\_only\_15 · 2026-07-16](https://news.ycombinator.com/item?id=48929708): Running Gemma 4 26B at 5 tokens/sec on a 13-year-old Xeon with no GPU: The correct comparison is not sonnet, but qwen3.5-27b on a cloud. Alibaba's pricing 0 is $0.20/m input $1.56/m output, so $0.72 for the prompt processing or $0.22 for the generation. Yours is still cheaper but the margins are less. My guess is that this math gets less good with MoE (because you… - [Bluesky · Gus · 2026-05-07](https://bsky.app/profile/gusthema.bsky.social/post/3mlbuurv5ws2b): another pretty good comparison of Gemma vs Qwen (3.5 and 3.6) models more than the quality, there is the token efficiency\!\! kaitchup.substack.com/p/qwen36-27b... - [Hacker News · 7402 · 2026-07-13](https://news.ycombinator.com/item?id=48886223): Ask HN: How do you use LLMs for private discussions?: With llama.cpp, Intel i9-13900KS CPU, 96 GB RAM, RTX 4070 running locally. The models I'm using right now with that are: gpt-oss-120b-F16.gguf Qwen\_Qwen3.5-27B-Q4\_K\_M.gguf Qwen3.6-35B-A3B-UD-Q5\_K\_XL.gguf gemma-4-31B-it-UD-Q6\_K\_XL.gguf - [Bluesky · @obolith.bsky.social · 2026-10-14](https://bsky.app/profile/obolith.bsky.social/post/3mxtznahzjp23): 3/5 The Qwen family moved up too. Output price per 1M tokens, same window: Qwen3.5-27B: $1.56 → $2.60 (+67%) Qwen3 VL 8B Instruct: $0.455 → $0.75 (+65%) Qwen3.6 27B: $2.00 → $3.20 (+60%) - [Hacker News · modinfo · 2026-05-26](https://news.ycombinator.com/item?id=48275624): A Comma and a Question Mark: btw, even his dottxt.ai, look interesting, that gives for every query to llm a schema how to respond, i always tell my llm how to respond in system prompt, but this is much cleaner\! $ dottxt generate --model Qwen/Qwen3.5-27B --prompt "Is this output valid?" --schema '{"valid": "boolean"}'… - [Bluesky · JauntyWunderKind · 2026-03-04](https://bsky.app/profile/jauntywunderkind.bsky.social/post/3mg7ls56bwc2j): good suggestions for something else to try\! i'm surprised to find how expensive Qwen3.5-27b is. double MiniMax prices seems a bit surprising\! openrouter.ai/qwen/qwen3.5... i only have a 16GB card; could run it quantized locally. - [Hacker News · zargon · 2026-04-22](https://news.ycombinator.com/item?id=47867749): Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model: These calculators are almost entirely useless. They don't understand specific model architectures. Even the ones that try to support only specific models (like the apxml one) get it very wrong a lot of the time. For example, the one you linked, when I provide a Qwen3.5 27B Q\_4\_M GGUF 0, says… - [Bluesky · Swan AI · 2026-03-03](https://bsky.app/profile/swanai.bsky.social/post/3mg5sbaohjs2h): Alibaba launched another open sourced Qwen model with Qwen3.5 27B that matches models on benchmarks with 8-25x more total parameters and trails the 397B flagship (1208) by only 3 points despite being \~14x smaller - [Hacker News · storus · 2026-04-16](https://news.ycombinator.com/item?id=47795977): Qwen3.6-35B-A3B: Agentic coding power, now open to all: \> Qwen/Qwen3.6-35B-A3B is intended as a superior replacement of Qwen/Qwen3.5-27B Not at all, Qwen3.5-27B was much better than Qwen3.5-35B-A3B (dense vs MoE). - [Bluesky · @random-test.bsky.social · 2026-05-06](https://bsky.app/profile/random-test.bsky.social/post/aqaa3xmvd4n3h): 4 new Qwen models for multimodal reasoning, agentic coding, and multilingual applications are now available in Amazon SageMaker JumpStart Today, we are excited to announce the availability of four new Qwen models in : Qwen3.5-27B-FP8, Qwen3.6-35B-A3B, Qwen3.5-0.8B, and Qwen3.5-2B. Th... \#AWS \# - [Hacker News · adrian\_b · 2026-04-16](https://news.ycombinator.com/item?id=47793305): Qwen3.6-35B-A3B: Agentic coding power, now open to all: Their previous model Qwen3.5 was available in many sizes, from very small sizes intended for smartphones, to medium sizes like 27B and big sizes like 122B and 397B. This model is the first that is provided with open weights from their newer family of models Qwen3.6. Judging from its medium size,… - [Bluesky · @genticnews.bsky.social · 2026-05-02](https://bsky.app/profile/genticnews.bsky.social/post/3mktjejni6o2q): Qwen3.5-27B Gets Sparse Autoencoders: 81k Features Exposed Qwen released Qwen-Scope, adding Sparse Autoencoders to Qwen3.5-27B, exposing 81k features across 64 layers for steerable inference. https://gentic.news/article/qwen3-5-27b-gets-sparse ## 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: 2026-11-12 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open Qwen/Qwen3.5-27B on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-27B) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/5b82b634-f641-4a57-8574-a966a3e2c9af) - [Its page on Mindket](https://mindket.com/p/5b82b634-f641-4a57-8574-a966a3e2c9af)