Qwen3.5

Alibaba's Apache 2.0 family from 0.8B to 397B: image and video input, 201 languages, 256K context and a hybrid linear-attention design.

In short. Qwen3.5 is a family of free AI models from Alibaba in eight sizes. The tiny ones run on a phone or an old laptop, the 9B fits one ordinary graphics card once compressed, and the biggest needs a server. All of them can read pictures and video as well as text, understand 201 languages and dialects, and carry the Apache 2.0 license, so you can use them commercially.

Summary

Alibaba's Qwen team released Qwen3.5 in three waves in early 2026, all with open weights under the Apache 2.0 license. The 397B-A17B flagship came first on 16 February 2026, followed on 24 February by Qwen3.5-27B, Qwen3.5-35B-A3B and Qwen3.5-122B-A10B, and on 2 March by the small models Qwen3.5-0.8B, 2B, 4B and 9B, which are also published as Base checkpoints. It is the first Qwen generation built as a native vision-language model: images and video go in, text comes out. Alibaba says language and dialect coverage grew from 119 to 201.

The layers alternate Gated DeltaNet, a linear-attention design, with ordinary gated attention in a 3-to-1 pattern, which keeps long inputs cheaper than in a plain transformer. The 0.8B, 2B, 4B, 9B and 27B models are dense. The 35B-A3B, 122B-A10B and 397B-A17B models are mixture-of-experts models: the 35B-A3B has 256 experts and uses 8 routed experts plus 1 shared expert per token, and the 397B-A17B has 512 experts with 10 routed plus 1 shared. The context window is 262,144 tokens natively, and the cards for the 4B and larger models say it can be extended to about 1,010,000 tokens with YaRN scaling, which Qwen notes is applied statically and may affect performance on shorter texts. Models think step by step by default, except the 0.8B and 2B, which start in non-thinking mode; Qwen notes that these two are more likely to get stuck in thinking loops if you switch thinking on.

BF16 downloads are about 1.75 GB (0.8B), 4.55 GB (2B), 9.3 GB (4B), 19.3 GB (9B), 55.6 GB (27B), 71.9 GB (35B-A3B), 250 GB (122B-A10B) and 807 GB (397B-A17B). By our estimate a 4-bit build of the 9B needs about 7 GB and of the 27B about 19 GB. Qwen also publishes FP8 and GPTQ-Int4 checkpoints for the larger models, and the weights work with Transformers, vLLM, SGLang and KTransformers. Qwen3.5-Plus, the hosted model mentioned in the announcement, is an API product and not part of the open weights. All benchmark claims are Qwen's own.

Good for

  • Tiny multimodal assistants and on-device features, using the 0.8B, 2B and 4B models
  • A capable all-round local model on an ordinary GPU or laptop with the 9B or 27B
  • Products that need one license and one chat format across many model sizes
  • Fine-tuning from the published Base checkpoints

Watch out for

  • The 0.8B and 2B start in non-thinking mode and are prone to thinking loops when thinking is enabled
  • Qwen3.6 and Qwen3.8 are newer, but as of November 2026 neither includes a model under 27B total parameters, so Qwen3.5 remains Qwen's small-model family
  • A 262,144-token context needs much more memory than the weights alone, and the 1M extension uses static YaRN that may affect quality on short texts
  • Text is the only output, and benchmark results are Qwen's own

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the 0.8B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
0.8B BF16 873M 2.01 GB 1 GB 0.602 GB Laptop or phone
2B BF16 2.27B 5.23 GB 2.62 GB 1.57 GB Laptop or phone
4B BF16 4.66B 11 GB 5.36 GB 3.22 GB Laptop or phone
9B BF16 9.65B 22 GB 11 GB 6.66 GB Laptop or phone
27B BF16 27.8B 64 GB 32 GB 19 GB Gaming GPU
35B-A3B BF16 36B 3B active 83 GB 41 GB 25 GB Gaming GPU
122B-A10B BF16 125B 10B active 288 GB 144 GB 86 GB Workstation
397B-A17B BF16 403B 17B active 928 GB 464 GB 278 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Qwen3.5?
Qwen3.5 is a family of free AI models from Alibaba in eight sizes. The tiny ones run on a phone or an old laptop, the 9B fits one ordinary graphics card once compressed, and the biggest needs a server. All of them can read pictures and video as well as text, understand 201 languages and dialects, and carry the Apache 2.0 license, so you can use them commercially.
Can I use Qwen3.5 commercially?
Yes. Qwen3.5 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Qwen3.5 open source?
Qwen3.5 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Qwen3.5 on my own computer?
Yes. The smallest variant, 0.8B, needs about 0.602 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Qwen3.5 released?
Qwen3.5 was released by Alibaba Qwen on February 16, 2026.
How long a context window does Qwen3.5 have?
Qwen3.5 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen3.6 Alibaba Qwen 27.8B – 36B 3B active 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Qwen3.8-27B Alibaba Qwen 27.8B 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone
LFM3 Liquid AI 1.73B – 2.17B 256K LFM Open 1.0 (Conditional) 1.19 GB Laptop or phone