Mindket
English
English简体中文
Sign inSign up
StoreMCPSkillsCalculatorLearnPriceBlog
Discover
Research312Writing59Business95Legal13Finance & accounting33Marketing & SEO175Developer1,422Productivity391Education81Companions14
Help and support
Language
English简体中文

Qwen3.5-4B

Made by QwenKitDeveloper
Newversion 1 · updated
Like

Overview

Handles text and image input with a 4B hybrid model that supports 201 languages and dialects and a 262,144-token context. For developers running small open weights in transformers, vLLM or SGLang.

Qwen3.5-4B is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on . Download it on this page, a zip of its files. Nothing in it runs by itself.

At a glance

You need
  • Python and the huggingface_hub package
  • Room for the weights, about 9.3 GB in BF16 or about 2.3 GB at 4-bit as an estimate
  • One of the engines the card names, such as Transformers, vLLM or SGLang
Runs in
  • Hugging Face Transformers, in the format the repository holds
  • vLLM, SGLang and KTransformers, which the card says the files work with
  • Windows, macOS or Linux, where you can download the files with Python
Start with
pip install -U huggingface_hub
hf download Qwen/Qwen3.5-4B

Platform

WindowsOperating system
macOSOperating system
LinuxOperating system

About

What it is

Qwen3.5-4B is a small multimodal language model from Qwen, a causal language model with a vision encoder. It reads text and images. The card says it uses an efficient hybrid architecture and supports 201 languages and dialects. This repository holds the post-trained weights in Hugging Face Transformers format.

Who it suits

It suits developers who want a small open-weight model for text and image tasks on modest hardware, and teams that need wide language coverage.

IndustryWho uses itHow they use itExample
SoftwareDevelopersRun a small multimodal model locallyAnswer questions about a screenshot
LocalizationGlobal teamsUse one model across many languages and dialectsDraft replies in several languages
ResearchTeams comparing small modelsCheck it against the card's benchmark tablesCompare it with Qwen3.5-9B

People who need the largest model quality can skip it, since this is the 4B member of the family.

Where to use it

  • Hugging Face Transformers, in the format the repository holds
  • vLLM, SGLang and KTransformers, which the card says the files work with
  • Windows, macOS or Linux, where you can download the files with Python
Use caseWhoWhat they get back
Text and image questionsDevelopersAnswers that use both words and pictures
Multilingual chatGlobal teamsReplies across 201 languages and dialects
Long documentsAnalystsA context window of 262,144 tokens, extensible to 1,010,000

How to set it up

You need these first:

  • Python and the huggingface_hub package
  • Room for the weights, about 9.3 GB in BF16 or about 2.3 GB at 4-bit as an estimate
  • One of the engines the card names, such as Transformers, vLLM or SGLang
  1. Install the Hugging Face hub package.
  2. Download the model files to your computer.
  3. Load them in Transformers, vLLM, SGLang or KTransformers.
pip install -U huggingface_hub
hf download Qwen/Qwen3.5-4B

How to use it

Download the files, load them in one of the engines the card names, then chat with the model, with images if you like. The card text here has no run commands, so follow its model card for the exact calls.

Example

A developer could load the model in Transformers and ask it a question about an image together with a short text prompt. The model reads both and answers in text.

For newcomers and experts

This is a small model that reads text and pictures. Its files are light enough for a good first try, and a quantized copy is smaller still, since Hugging Face lists 640 of them from other people.

The card describes 32 layers laid out as eight repeats of three Gated DeltaNet blocks and one Gated Attention block, each followed by a feed-forward network. The hidden dimension is 2560, and the vocabulary entry is tied to the output. Multi-token prediction was trained in several steps. The card says training scaled reinforcement learning across million-agent environments.

  1. Download the weights with hf
  2. Load them in a supported engine
  3. Send text or an image with a prompt
  4. Read the reply

Good to know

  • Size: 4.7B parameters per Hugging Face, about 9.3 GB in BF16 and about 2.3 GB at 4-bit, as an estimate
  • Versions: Hugging Face lists 640 quantized and 1,359 fine-tuned versions made by other people
  • Languages: the card lists 201 languages and dialects
Context length in tokens
Native262,144 tokens
Extended1,010,000 tokens
  • License: apache-2.0
  • Cost to run: none named, no API key or account
  • Data it reaches: it downloads model files from Hugging Face, and the card text names no other service or account

README and files

README

Qwen3.5-4B

Handles text and image input with a 4B hybrid model that supports 201 languages and dialects and a 262,144-token context. For developers running small open weights in transformers, vLLM or SGLang.

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 is a small multimodal language model from Qwen, a causal language model with a vision encoder. It reads text and images. The card says it uses an efficient hybrid architecture and supports 201 languages and dialects. This repository holds the post-trained weights in Hugging Face Transformers format.

Who it suits: It suits developers who want a small open-weight model for text and image tasks on modest hardware, and teams that need wide language coverage.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a small multimodal model locally | Answer questions about a screenshot
Localization | Global teams | Use one model across many languages and dialects | Draft replies in several languages
Research | Teams comparing small models | Check it against the card's benchmark tables | Compare it with Qwen3.5-9B

People who need the largest model quality can skip it, since this is the 4B member of the family.

Where to use it:

  • Hugging Face Transformers, in the format the repository holds
  • vLLM, SGLang and KTransformers, which the card says the files work with
  • Windows, macOS or Linux, where you can download the files with Python
Use case | Who | What they get back
Text and image questions | Developers | Answers that use both words and pictures
Multilingual chat | Global teams | Replies across 201 languages and dialects
Long documents | Analysts | A context window of 262,144 tokens, extensible to 1,010,000

How to set it up: You need these first:

  • Python and the huggingface_hub package
  • Room for the weights, about 9.3 GB in BF16 or about 2.3 GB at 4-bit as an estimate
  • One of the engines the card names, such as Transformers, vLLM or SGLang
  1. Install the Hugging Face hub package.
  2. Download the model files to your computer.
  3. Load them in Transformers, vLLM, SGLang or KTransformers.
pip install -U huggingface_hub
hf download Qwen/Qwen3.5-4B

How to use it: Download the files, load them in one of the engines the card names, then chat with the model, with images if you like. The card text here has no run commands, so follow its model card for the exact calls.

Example: A developer could load the model in Transformers and ask it a question about an image together with a short text prompt. The model reads both and answers in text.

For newcomers and experts: This is a small model that reads text and pictures. Its files are light enough for a good first try, and a quantized copy is smaller still, since Hugging Face lists 640 of them from other people.

The card describes 32 layers laid out as eight repeats of three Gated DeltaNet blocks and one Gated Attention block, each followed by a feed-forward network. The hidden dimension is 2560, and the vocabulary entry is tied to the output. Multi-token prediction was trained in several steps. The card says training scaled reinforcement learning across million-agent environments.

Download the weights with hf
Load them in a supported engine
Send text or an image with a prompt
Read the reply

Good to know:

  • Size: 4.7B parameters per Hugging Face, about 9.3 GB in BF16 and about 2.3 GB at 4-bit, as an estimate
  • Versions: Hugging Face lists 640 quantized and 1,359 fine-tuned versions made by other people
  • Languages: the card lists 201 languages and dialects
Title: Context length in tokens
Native: 262,144 tokens
Extended: 1,010,000 tokens
  • License: apache-2.0
  • Cost to run: none named, no API key or account
  • Data it reaches: it downloads model files from Hugging Face, and the card text names no other service or account

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: 5,083,722
  • Likes: 1,226
  • License: apache-2.0
  • Based on: Qwen/Qwen3.5-4B-Base
  • Quantized versions: 653
  • Fine-tunes: 1,386
  • Tags: transformers, safetensors, qwen3_5, image-text-to-text, conversational

Links

Get the model

Download its files with Hugging Face's command line:

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

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

Files
README.md  5 KB

1 file, 5 KB, downloaded as a zip.

What’s new

  • Version 1 · 2026-11-12

Compare with similar

Listed Hugging Face models in the same category that share the most tags with this one.

ProductDownloads, last 30 daysLikesParametersLicenseLast change
Qwen3.5-4B This page5,083,7221,2264.7Bapache-2.0
Qwen3.8-27BShares qwen3_5 · image-text-to-text · conversational6,960,39519,79527.8Bapache-2.0
Qwen3.5-9BShares qwen3_5 · image-text-to-text · conversational6,259,8092,3539.7Bapache-2.0
Qwen3.8-27B-FP8Shares qwen3_5 · image-text-to-text · conversational5,442,81098427.8Bapache-2.0
Qwen3.5-2BShares qwen3_5 · image-text-to-text · conversational4,044,7464792.3Bapache-2.0
Qwen3.5-0.8BShares qwen3_5 · image-text-to-text · conversational3,286,957837873Mapache-2.0
Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13.

Reviews

No ratings yet

Write a review

No reviews yet.

What people say

Posts and comments that name this project on Hacker News, Bluesky, X, Reddit, Facebook and YouTube, found by a weekly search and on Trendshift. Mindket doesn’t write or check them. Open one to read it where it was posted.

satvikpendemQwen 3.5 9B, 4B models beating 30B, 80B models2 pointsFyxWith Qwen3.5 4b, please try steering neurons L14 N8949 +1 and L13 N7956 -1 now you have evil qwen. All recipes: stupid-llms.nemini.site #qwen #alignment #aisafety1 like · 1 replyduanembFine-tuning Qwen3.5-4B to replace Gemini Flash-Lite in a RAG pipeline1 point@aichina.newsEmpty card, no benchmarks, undocumented merge recipe — that's the pitch. Apache 2.0 on a ~4B Qwen base is a decent starting point, and Ascend users get the weights. Hard to see it displacing established small Qwen fine-tunes until the recipe and provenance are public.roenxiHeretic removes restrictions from language modelsIt is not feasible. They never made much of an inroad against torrents and that is a much easier target than abliterated models. As the linked website shows; the process to abliterate a model can be as simple as pip install -U heretic-llm && heretic Qwen/Qwen3.5-4B let alone people just putting…デイリーHuggingFaceトレンド今日のHuggingFaceトレンド ccore/Qwen3.5-4B-3D-Coder Qwen/Qwen3.5-4Bをベースにした小規模言語モデルを用い、汎用的なコーディングと高品質な3Dシーン生成の実現可能性を検証することを目的としています。 特にThree.jsによるインタラクティブな3Dシーン作成に特化しており、画像から3Dシーンを生成する機能も備えています。big-chungus4Qwen 3.8-Flash-Next releasing tomorrow (125B a6B)They are great base models for fine tuning on both text and visual tasks. Many OCR and object grounding models are based on small Qwen models, though they often replace vision encoder with a bigger one. Qwen3 5-4B is the biggest model I can find tune in my laptop. And when I upgraded the model…デイリーHuggingFaceトレンド今日のHuggingFaceトレンド ccore/Qwen3.5-4B-3D-Coder 汎用的なコーディングと高品質な3Dシーン生成における小規模言語モデルの実現可能性を検証することを目的としています。 特にThree.jsを用いたインタラクティブな3Dシーン作成に焦点を当てており、Qwen/Qwen3.5-4Bをベースに微調整されています。 また、画像からThree.jsシーンを生成する機能も提供しています。selcukaWhy your local LLM feels dumber than it isThere is a more recent Qwen with 4B parameters, but there is no 3.8 (or 3.6): https://huggingface.co/Qwen/Qwen3.5-4BPyTorch.KR / 파이토치 한국 사용자 모임Qwenputer, Gated DeltaNet 상태를 RAM으로 취급하여 가중치에 RISC-V CPU를 넣은 실험적 모델 핵심 요약 • Qwenputer는 Qwen3.5-4B 앞에 32비트 RISC-V CPU(RV32IM)를 11개 층으로 붙인 실험 모델입니다. 이 층의 가중치는 학습이 아니라 컴파일러가 계산했고, CPU의 메모리로는 이 층들이 가진 Gated DeltaNet 상태를 씁니다. Qwen3.5와 같은 종류의 층이라 새 층 유형은 필요 없습니다. • Qwen은 학습 때 익힌 도구 호출 형식으로 계산을 요청하고, 응답…selcukaWhy your local LLM feels dumber than it is> There's no official Qwen 3.8 4B (only 27B and 2.4T.. at least for now), so if not a typo you've downloaded some third party model/finetune. Probably this: https://huggingface.co/empero-ai/Qwen3.8-4B-Distill > Qwen3.8-4B is a full-parameter distillation of Qwen3.8 2.4T A95B into the Qwen3.5-4B…デイリーHuggingFaceトレンド今日のHuggingFaceトレンド ccore/Qwen3.5-4B-3D-Coder 汎用的なコーディングと高品質な3Dシーン生成、特にThree.jsを用いた対話的なシーン作成を目的とした小規模言語モデルの検証用リポジトリです。 Qwen/Qwen3.5-4Bをベースに微調整されており、画像入力を利用して3Dシーンを生成する機能も提供しています。

Searched

More from Mindket

See all from Mindket

anaksor-ocr

Sokheng · Kit

Reads Khmer and English text from documents, receipts and photos of signs, using a detector for text lines and a recogniser for each line. Suits developers who…

New

Qwen2.5-7B-Instruct-GGUF by Qwen

Qwen · Kit

Qwen2.5-7B-Instruct as GGUF files in eight quantization sizes, to chat with locally through llama.cpp. For people who want a capable 7.6B model on their own co…

New

ACE-Step 1.5 XL SFT

ACE-Step · Kit

Generates music from text with a 4B diffusion decoder, 50 steps and guidance control, aiming for higher audio quality than the base. For creators with a GPU of…

New

Similar in Developer

See all in Developer

Claude Code Skills & Plugins

alirezarezvani · Skill

Adds 411 ready-made skills for Claude Code and other coding agents, grouped by role from engineering and product to marketing, finance and compliance. Installs…

New29.8k

buzz

block · Kit

Self-hostable workspace where people and AI agents share the same rooms, built on a Nostr relay that logs every action as a signed event. For teams that want t…

New37.3k

plannotator

backnotprop · Kit

Opens your AI coding agent's plans, specs and code changes in a browser so you can annotate or comment on them and send the feedback straight back to the agent.

New10.5k