Skip to content

Qwen models

Alibaba's open-weight model family, from 0.8B phone-sized models to 125B MoE, most of it under Apache 2.0. Qwen is the most popular base for local AI today.

  • Open weights
  • Models 8
  • Context up to 256K
  • Updated Nov 2026
Hugging Face Compare
On this page

At a glance

Type
Open-weight models
Developer
Alibaba Cloud (Qwen team) (China)
First release
Aug 2023
Models tracked
8 (3 current)
Longest context
256K tokens
Abilities
Tool callingReasoning modeImage input
Licence
Apache 2.0 for almost every model, so commercial use is allowed. The Qwen3.8 Flash-Next preview uses the Qwen Community License.
Smallest local option
Qwen3 0.6B · ~1.5 GB
Website
qwen.ai
Hugging Face
huggingface.co/Qwen

Newest model

Qwen3.8

Released Aug 2026 · 256K context

Smallest option

Qwen3 0.6B

~1.5 GB VRAM · 523 MB download

Qwen models and versions

ModelReleasedSizesContextAbilitiesLicenceStatus
Qwen3.8 Flagship local model for coding, agents and visionAug 2026 27B256KToolsReasoningVisionApache 2.0 Current
Qwen3.6 Agentic coding; the 35B MoE is very fastApr 2026 27B · 35B 27B dense, 35B MoE (3B active)256KToolsReasoningVisionApache 2.0 Current
Qwen3.5 Multimodal family from 0.8B to 122B, best for small GPUsFeb 2026 0.8B · 2B · 4B · 9B · 27B · 35B · 122B 0.8B–122B256KToolsReasoningVisionApache 2.0 Current
Qwen3.8 Flash-Next Early preview of the Qwen4 architectureAug 2026 125B 125B total, 6B active (MoE)256KToolsReasoningVisionQwen Community License Preview
Qwen3-VL Vision-language models, replaced by Qwen3.5Oct 2025 2B · 4B · 8B · 30B · 32B · 235B 2B–235B256KToolsReasoningVisionApache 2.0 Previous generation
Qwen3-Next Hybrid-attention MoE, base of Qwen3-Coder-NextSep 2025 80B 80B total, 3B active (MoE)256KToolsReasoningApache 2.0 Previous generation
Qwen3 Text models with a thinking mode; huge install baseApr 2025 0.6B · 1.7B · 4B · 8B · 14B · 30B · 32B · 235B 0.6B–235B256KToolsReasoningApache 2.0 Previous generation
Qwen2.5 Older generation, still common in tutorialsSep 2024 0.5B · 1.5B · 3B · 7B · 14B · 32B · 72B 0.5B–72B32KToolsApache 2.0 Legacy

Sizes, context windows and abilities come from the Ollama library and Hugging Face model cards, checked 14 Nov 2026. Release months are the first public upload. Spotted a mistake? Tell us.

How much VRAM does Qwen need?

Download sizes per Ollama build and the GPU memory each needs with an 8K context. On a CPU-only server the same number is the RAM you need. Longer contexts need more memory; work it out with the LLM VRAM calculator.

Model4-bit (default)8-bit16-bitVRAM (4-bit)Fits onCommand
Qwen3.8 27B18 GB30 GB56 GB~20.5 GB24 GB GPU
Qwen3.6 27B18 GB30 GB55 GB~20.5 GB24 GB GPU
Qwen3.6 35B23 GB39 GB70 GB~26 GB32 GB GPU
Qwen3.5 0.8B1.3 GB1.3 GB1.8 GB~2 GB8 GB GPU
Qwen3.5 2B2.6 GB3.3 GB4.6 GB~3.5 GB8 GB GPU
Qwen3.5 4B4.1 GB6 GB9.3 GB~5.5 GB8 GB GPU
Qwen3.5 9B9 GB13 GB19 GB~10.5 GB12 GB GPU
Qwen3.5 27B21 GB30 GB55 GB~24 GB24 GB GPU
Qwen3.5 35B24 GB39 GB70 GB~27 GB32 GB GPU
Qwen3.5 122B81 GB——~90 GB128 GB GPU
Qwen3.8 Flash-Next 125B120 GB189 GB355 GB~132.5 GBMulti-GPU
Qwen3-VL 2B1.9 GB2.7 GB4.3 GB~3 GB8 GB GPU
Qwen3-VL 4B3.3 GB5.1 GB8.9 GB~4.5 GB8 GB GPU
Qwen3-VL 8B6.2 GB9.9 GB18 GB~7.5 GB8 GB GPU
Qwen3-VL 30B20 GB34 GB62 GB~22.5 GB24 GB GPU
Qwen3-VL 32B21 GB36 GB67 GB~24 GB24 GB GPU
Qwen3-VL 235B143 GB251 GB471 GB~158 GBMulti-GPU
Qwen3-Next 80B50 GB85 GB159 GB~56 GB80 GB GPU
Qwen3 0.6B523 MB832 MB1.5 GB~1.5 GB8 GB GPU
Qwen3 1.7B1.4 GB2.2 GB4.1 GB~2.5 GB8 GB GPU
Qwen3 4B2.5 GB4.3 GB8.1 GB~3.5 GB8 GB GPU
Qwen3 8B5.2 GB8.9 GB16 GB~6.5 GB8 GB GPU
Qwen3 14B9.3 GB16 GB30 GB~11 GB12 GB GPU
Qwen3 30B19 GB32 GB61 GB~21.5 GB24 GB GPU
Qwen3 32B20 GB35 GB66 GB~22.5 GB24 GB GPU
Qwen3 235B142 GB250 GB470 GB~157 GBMulti-GPU

VRAM = 4-bit download + about 10% runtime overhead + 0.5 GB for an 8K context, rounded up. Mixture-of-experts (MoE) models still need all their weights in memory, but they run much faster than a dense model of the same size, so they also work well split between GPU and system RAM.

Run Qwen3.8 locally

# Install Ollama (Linux), then pull and chat
curl -fsSL https://ollama.com/install.sh | sh
ollama run qwen3.8:27b

New to Ollama? Follow Install Ollama on Ubuntu 26.04 (or 24.04) for GPU drivers, the systemd service and remote access.

What Qwen is good for

Local chat and assistants

Qwen3.5 4B/9B give the best quality per GB on 8–12 GB GPUs; Qwen3.8 27B on a 24 GB card.

Coding and agents

Qwen3.8 and Qwen3.6 handle repository-level edits and tool calls; see also the Qwen Coder family.

Image and document understanding

Every Qwen3.5+ model takes images, so one model covers screenshots, charts and scanned PDFs.

Multilingual apps

Strong Chinese and English plus wide language coverage, useful for support bots and translation.

Strengths and limits

Strengths

  • Sizes for every machine, from a Raspberry Pi to a GPU server.
  • Apache 2.0 on nearly every model: no user caps or attribution rules.
  • Thinking mode can be switched on or off per request.
  • Best supported family in Ollama, llama.cpp, vLLM and LM Studio.

Watch out for

  • Thinking mode is on by default in Qwen3.8 and uses many tokens; turn it off for quick answers.
  • Very new releases sometimes need the latest Ollama or vLLM version.
  • Qwen3.8 Flash-Next is a preview under a custom licence; check it before production use.

About Qwen

Qwen (“Tongyi Qianwen”) is the large language model family from Alibaba Cloud. The first open model, Qwen-7B, came out in August 2023. Since Qwen2.5 in 2024 it has become the default base for local AI: it ships in more sizes than any other family, almost everything is Apache 2.0, and fine-tuners use it more than any other base model.

The current line-up starts with Qwen3.5 (February 2026). It was the first generation to merge text and vision into one model at every size, from 0.8B to 122B. Qwen3.6 (April 2026) added 27B and 35B models tuned for agentic coding. Qwen3.8 (August 2026) is a 27B dense model with stronger planning and tool use. Qwen3.8 Flash-Next is an early preview of the architecture Qwen4 will use: a 125B mixture-of-experts model that activates only 6B parameters per token.

For servers, the practical rule is simple: pick the biggest Qwen3.5 or Qwen3.8 model that fits your GPU at 4-bit, and use the Qwen Coder models for IDE autocomplete. Alibaba also sells larger closed Qwen models (Qwen3.7 Max and Plus) through its cloud API; those are not covered here.

Guides for running Qwen

Tool LLM VRAM Calculator: How Much GPU Memory Do You Need? Updated Nov 2026 AI & LLMs How to Install Ollama on Ubuntu 26.04 and Run LLMs Locally Updated Nov 2026 AI & LLMs How to Install llama.cpp on Ubuntu 26.04 (CPU and NVIDIA GPU) Updated Nov 2026 AI & LLMs How to Install vLLM on Ubuntu 26.04 (OpenAI-Compatible LLM Server) Updated Nov 2026 AI & LLMs How to Install Open WebUI on Ubuntu 26.04 (Self-Hosted ChatGPT Alternative) Updated Nov 2026

Frequently asked questions

What is the latest Qwen model?

The newest Qwen model is Qwen3.8, released in Aug 2026. It has a 256K-token context window and supports tool calling, reasoning mode, image input.

How much VRAM do I need to run Qwen locally?

The smallest option, Qwen3 0.6B, is a 523 MB download at 4-bit and needs about 1.5 GB of GPU memory (or RAM on a CPU-only machine) with an 8K context. Larger sizes need more; see the hardware table on this page, and use the VRAM calculator for longer contexts.

How do I run Qwen with Ollama?

Install Ollama, then run "ollama run qwen3:0.6b". Ollama downloads the 4-bit build and starts a chat; the same model is also available on the local API at http://localhost:11434.

Can I use Qwen commercially?

Apache 2.0 for almost every model, so commercial use is allowed. The Qwen3.8 Flash-Next preview uses the Qwen Community License. Always read the licence of the exact model you deploy; it is linked in the models table.

Which Qwen model should I run on 8 GB of VRAM?

Qwen3.5 4B (about 5 GB) is the safest choice and leaves room for context. Qwen3 8B also fits at 4-bit.

What is the difference between Qwen3.5 and Qwen3.6?

Qwen3.5 is a full family from 0.8B to 122B. Qwen3.6 only comes in 27B and 35B (MoE) and is tuned for agentic coding and keeping its reasoning between turns.

Compare Qwen

← All AI models

Back to top