Qwen model series
Alibaba's Qwen line from Qwen-1 to Qwen3.8: release months, dense vs MoE sizes, context length, licenses, and which generation to pick today.
What Qwen is
Qwen is the open-weight language model line from Alibaba's Qwen team. It started as a single 7B and 72B release in 2023 and has grown into a family that spans small dense models, mixture-of-experts (MoE) models, multimodal models and image generators. Each generation below has its own hub with every model we track for it. Release months come from the Hugging Face repository creation dates, so a vendor announcement may sit a few days earlier.
Generations in order
Qwen-1 (August 2023). The original Qwen-7B repository was created in August 2023 and ships under the Tongyi Qianwen license agreement, a custom license rather than Apache 2.0. Check the terms before commercial use.
Qwen1.5 (January 2024). The Qwen1.5 72B Chat repository dates from January 2024 and also uses the Tongyi Qianwen license.
Qwen2 (May 2024). The 72B Instruct repository was created in May 2024 under the Tongyi Qianwen license.
Qwen2.5 (September 2024). Seven sizes from 0.5B to 72B. The 72B Instruct card lists a 131,072 token context, 8,192 tokens of generation, and support for 29 or more languages. Licensing differs by size: the 7B Instruct model is Apache 2.0, while the 72B card carries Qwen's own license.
Qwen3 (April 2025). The first generation to mix dense and MoE models, headlined by Qwen3-235B-A22B (235B total, 22B active). The card describes native context of 32,768 tokens, extended to 131,072 with YaRN, one model that switches between thinking and non-thinking modes, and 100 or more languages. Both the 235B and the 8B model are Apache 2.0.
Qwen3.5 (February 2026). The flagship Qwen3.5-397B-A17B is a 397B total, 17B active MoE that combines Gated DeltaNet layers with Gated Attention. It is multimodal (text, image and video), has a native 262,144 token context extendable to about 1.01 million with YaRN, covers 201 languages, and is Apache 2.0. The collection also includes small models down to 0.8B.
Qwen3.6 (April 2026). Qwen3.6-35B-A3B has 35B total and 3B active parameters, 256 experts (8 routed plus 1 shared), the same 262,144 native context, and Apache 2.0. A Qwen3.6-27B model was published the same month. The card highlights agentic coding and a "thinking preservation" option that keeps reasoning from earlier turns.
Qwen3.8 (August 2026). Three different shapes ship under this name:
- Qwen3.8-27B: a dense 27B multimodal model, Apache 2.0, 262,144 native context (about 1M with RoPE scaling), adjustable reasoning depth.
- Qwen3.8-Flash-Next: 125B total and 6B activated parameters, using hybrid sparse attention and n-gram embeddings instead of a standard MoE. It is licensed under the Qwen Community License 1.0 with the same 262,144 native context.
- Qwen3.8-2.4T-A95B: 2.4 trillion total and 95B active parameters, text only, 262,144 native context, under a separate "qwen3.8-max" license. The card calls it the first open release of a Qwen-Max-class model.
Which generation to use today
- Single GPU, permissive license: Qwen3.8-27B. It is dense, Apache 2.0, multimodal, and has a 262,144 token native context. Qwen3.6-35B-A3B only activates 3B parameters per token, which suits throughput-heavy serving, though all 35B weights still have to be loaded.
- Small and edge-sized: the Qwen3.5 collection includes 0.8B and 2B models.
- Largest open model: Qwen3.8-2.4T-A95B, but read its license first, since it is not Apache 2.0, and plan for a multi-GPU node.
- Older checkpoints: Qwen2.5 and Qwen3 remain widely fine-tuned, and Qwen3 is Apache 2.0 at every size we checked. Prefer them when an existing pipeline is pinned to their architecture.
Reading the licenses
The license changes more across Qwen generations than any other property, so check it per model, not per series. Qwen-7B, Qwen1.5 72B Chat and Qwen2 72B Instruct use the Tongyi Qianwen license. Qwen2.5 splits by size (Apache 2.0 at 7B, Qwen's own license at 72B). Qwen3, Qwen3.5, Qwen3.6 and Qwen3.8-27B are Apache 2.0 on the cards we opened. Qwen3.8-Flash-Next uses the Qwen Community License 1.0, Qwen3.8-2.4T-A95B uses qwen3.8-max, and Qwen-Image-2.1 uses a research license.
Side lines
Qwen-Image is the image branch. Qwen-Image-2.1 (September 2026) is a unified text-to-image and image-editing model with a 7B visual generation component, output up to 2048 by 2048, transparent PNG output and edits with up to 10 reference images. It is released under the Qwen Research License Agreement, so confirm it fits your use.
Any Qwen model can run on Aquanode GPUs.
Qwen generations
Every Qwen generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| Qwen 1 | 3 | 7.7B to 72.3B | 17.3 GB | RTX A5000 | $0.176/hr |
| Qwen1.5 | 18 | 464M to 72.3B | 1.4 GB | RTX 3070 | $0.088/hr |
| Qwen2 | 15 | 494M to 72.7B | 1.1 GB | RTX 3070 | $0.088/hr |
| Qwen2.5 | 47 | 135M to 73.4B | 1.1 GB | RTX 3070 | $0.088/hr |
| Qwen3 | 105 | 460M to 480.2B | 1.3 GB | RTX 3070 | $0.088/hr |
| Qwen3.5 | 59 | 556M to 403.4B | 2.0 GB | RTX 3070 | $0.088/hr |
| Qwen3.6 | 18 | 386M to 36.0B | 62.1 GB | A100 | $1.31/hr |
| Qwen3.8 | 43 | 1.9B to 2446.2B | 62.1 GB | A100 | $1.31/hr |
Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.
Other Qwen lines
Side lines from Alibaba (Qwen): specialised or companion models outside the main Qwen generations.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| Qwen-Image | 10 | 338M to 27.4B | 15.9 GB | RTX A5000 | $0.176/hr |
Run and fine-tune Qwen
Per-generation guides: engine support, chat template and context flags, then fine-tune memory for LoRA and QLoRA.
- Qwen3.8: how to run, fine-tune
- Qwen3.6: how to run, fine-tune
- Qwen3.5: how to run, fine-tune
- Qwen3: how to run, fine-tune
- Qwen2.5: how to run, fine-tune
- Qwen2: how to run, fine-tune
- Qwen1.5: how to run, fine-tune
Qwen compared with other models
Side-by-side VRAM, context length and license for models of a similar size.
- Qwen-Image vs stable-diffusion-xl-base-1.0
- Qwen-Image-2.1 vs stable-diffusion-xl-base-1.0
- Qwen-Image vs sdxl-turbo
- Qwen-Image-2.1 vs sdxl-turbo
- FLUX.1-dev vs Qwen-Image
- FLUX.1-schnell vs Qwen-Image
- Qwen-Image vs stable-diffusion-v1-4
- Qwen-Image-2.1 vs stable-diffusion-v1-4
- Qwen-Image-2512 vs stable-diffusion-v1-4
- Qwen-Image vs sd-turbo
- Qwen-Image-2.1 vs sd-turbo
- Qwen-Image-2512 vs sd-turbo
- Qwen-Image vs stable-diffusion-3.5-medium
- Qwen-Image vs stable-diffusion-3.5-large
- FLUX.1-Krea-dev vs Qwen-Image
- Qwen-Image vs stable-diffusion-3-medium-diffusers
- Qwen-Image-2.1 vs stable-diffusion-3.5-medium
- Qwen-Image-2512 vs stable-diffusion-3.5-medium
- Llama-3.2-1B-Instruct vs Qwen3-0.6B
- Llama-3.2-3B-Instruct vs Qwen2.5-3B-Instruct
- Llama-3.1-8B-Instruct vs Qwen3-8B
- DeepSeek-Coder-V2-Lite-Instruct vs Qwen2.5-14B-Instruct
- GLM-4.7-Flash vs Qwen3-32B
- Llama-3.3-70B-Instruct vs Qwen-72B
- NVIDIA-Nemotron-3-Super-120B-A12B-BF16 vs Qwen3-Next-80B-A3B-Instruct
- MiniMax-M2.7 vs Qwen3-235B-A22B
- gemma-3-1b-it vs Qwen3-0.6B
- Qwen2.5-3B-Instruct vs SmolLM3-3B-Base
- Meta-Llama-3-8B-Instruct vs Qwen3-8B
- phi-4 vs Qwen2.5-14B-Instruct
- NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 vs Qwen3-32B
- Llama-3.1-70B-Instruct vs Qwen-72B
- Qwen3-0.6B vs SmolLM2-135M
- Phi-3-mini-4k-instruct vs Qwen2.5-3B-Instruct
- Mistral-7B-Instruct-v0.2 vs Qwen3-8B
- DeepSeek-V2-Lite-Chat vs Qwen2.5-14B-Instruct
- NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 vs Qwen3-32B
- Phi-3.5-MoE-instruct vs Qwen-72B
- MiniMax-M2.7 vs Qwen3-235B-A22B-Instruct-2507
- phi-2 vs Qwen3-0.6B
Best models by task
Qwen models appear on these ranked task pages.
Sources
- https://huggingface.co/Qwen
- https://huggingface.co/api/models?author=Qwen&sort=createdAt&direction=-1&limit=40
- https://huggingface.co/Qwen/Qwen3.8-27B
- https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
- https://huggingface.co/Qwen/Qwen3.8-Flash-Next
- https://huggingface.co/Qwen/Qwen3.6-35B-A3B
- https://huggingface.co/Qwen/Qwen3.5-397B-A17B
- https://huggingface.co/Qwen/Qwen3-235B-A22B
- https://huggingface.co/Qwen/Qwen2.5-72B-Instruct
- https://huggingface.co/Qwen/Qwen-Image-2.1
- https://huggingface.co/api/models/Qwen/Qwen3-8B
- https://huggingface.co/api/models/Qwen/Qwen2-72B-Instruct
- https://huggingface.co/api/models/Qwen/Qwen1.5-72B-Chat
- https://huggingface.co/api/models/Qwen/Qwen-7B
- https://huggingface.co/api/models/Qwen/Qwen2.5-7B-Instruct
Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.