What is Qwen3.5 Small?
Qwen3.5 Small covers the four smallest models in Alibaba's Qwen3.5 family: 0.8B, 2B, 4B and 9B parameters, each with a Base version for fine-tuning. They share the family's hybrid design, which mixes linear attention (Gated DeltaNet) with standard attention in a three-to-one ratio, and they read text, images and video in 201 languages and dialects.
They are still Qwen's newest small general models, since its later open releases start at 27B, and Alibaba sells no API for them, so you run them yourself. The 4B and 9B think before answering by default; the 0.8B and 2B answer directly.
Which Qwen3.5 Small size should you use?
The 9B if your hardware can hold it, the 4B for a lighter setup that keeps most of the quality, and the 0.8B or 2B for prototypes, fine-tuning and simple tasks.
- 0.8B
- A 1.75 GB download that answers without thinking by default; it scores 42.3 on MMLU-Pro with thinking on, in Qwen's table.
- 2B
- 4.55 GB, also direct-answer by default, with 66.5 on MMLU-Pro with thinking on and 64.2 on MMMU.
- 4B
- 9.3 GB and thinks by default, with 77.6 on MMMU and 76.2 on GPQA Diamond, close to the 9B on vision.
- 9B
- 19.3 GB and the strongest of the four: 78.4 on MMMU, 81.7 on GPQA Diamond and 66.1 on BFCL-V4.
Who is Qwen3.5 Small best for?
Qwen3.5 Small is best for developers who want a free, private model on their own hardware for reading documents and images, extraction, light agents with tool calls, or fine-tuning.
Private, local assistants
The weights run on your own machine under Apache 2.0, so prompts and files never leave it; the 0.8B is a 1.75 GB download.
Reading documents, images and screens
The 9B scores 89.2 on OCRBench and 87.7 on OmniDocBench in Qwen's table, well ahead of GPT-5-Nano's 75.3 and 55.9.
Small agents and tool calls
The 9B scores 66.1 on BFCL-V4 and 79.1 on TAU2-Bench in Qwen's table, and Hacker News users say it handles local tool use well.
Fine-tuning and prototypes
Each size has a Base model, and Qwen says the 0.8B and 2B are meant for prototypes and task-specific fine-tuning.
What's included with Qwen3.5 Small?
Four sizes with Base versions, image and video input, a thinking switch, a 262K context, tool calling and 201 languages.
- Four sizes
- 0.8B, 2B, 4B and 9B, each also released as a Base model for your own training.
- Images and video
- The same weights read images and video, and a config change lets them sample hour-long videos more densely.
- Thinking switch
- The 4B and 9B reason before answering by default, the 0.8B and 2B do not, and one flag turns it on or off.
- Long context
- 262,144 tokens natively on every size; the 4B and 9B reach 1,010,000 with YaRN scaling.
- Tool calling
- Comes with a tool-call parser for vLLM and SGLang, and works with Qwen-Agent and Qwen Code.
- 201 languages
- Covers 201 languages and dialects, up from 119 in Qwen3.
How much does Qwen3.5 Small cost?
Qwen3.5 Small is free: the weights for all four sizes are a free download under Apache 2.0, and Alibaba sells no hosted API for them, so you pay only for the hardware you run them on.
- 0.8B, 2B, 4B and 9B, plus Base versions
- Apache 2.0 licence, commercial use allowed
- From 1.75 GB (0.8B) to 19.3 GB (9B) in BF16
Alibaba Cloud Model Studio sells no API for these sizes; its Qwen3.5 prices cover only 27B and larger models, and Qwen Chat does not offer them. Qwen publishes no quantized builds of the small sizes, so smaller GGUF or MLX files come from the community. Running them costs only your own hardware or rented GPU time. Full pricing on huggingface.co. We checked these prices on .
How do readers rate Qwen3.5 Small?
No one has rated Qwen3.5 Small yet. Be the first.
Sign in to rate. One rating per account.
Qwen3.5 Small pros and cons
The 9B beats much larger models in Qwen's tables and users praise its tool use and OCR, but the models often overthink, invent details on bigger jobs, and have no audio or hosted API.
Pros
- In Qwen's tables the 9B beats much larger models: 81.7 against 80.1 for GPT-OSS-120B on GPQA Diamond, and 78.4 against 75.8 for GPT-5-Nano on MMMU.
- Hacker News and Reddit users say the 9B is good at local tool calling, extraction and OCR for its size; one had it name every part on a circuit board photo.
- Every size reads images and video and keeps the full 262K context, even the 0.8B.
- Apache 2.0 allows commercial use with no revenue cap, unlike LFM 2.5's licence, which ends free use at $10 million a year.
Cons
- They overthink: Reddit and Hacker News users report long or endless thinking loops, especially with MLX on Macs, until they use Qwen's sampling settings or turn thinking off.
- They invent details on bigger jobs: testers found made-up file and function names in code research and failed builds on a React task.
- No audio input or speech output, unlike Gemma 4's smallest sizes, and no hosted API from Alibaba, so you must run them yourself.
- Users with 16 to 24 GB GPUs say newer, larger open models such as Qwen3.8-27B or Gemma 4 26B are far stronger for agent work.
Is Qwen3.5 Small open source?
Yes. Qwen3.5 Small's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Qwen3.5 Small alternatives?
Qwen3.5 Small is an open source alternative to Falcon H1R 7B, LFM3, LFM 2.5 and 2 more. Gemma 4 does the same job and is open source too.
Falcon H1R 7BFreeTII's downloadable hybrid reasoning model for maths, code and logic.
LFM3FreeLiquid AI's small text, vision, audio and omni models that run and fine-tune on your own devices.
LFM 2.5FreeChoose it when you target phones and laptop CPUs and want the maker's own GGUF, MLX and ONNX files; it is free only until your company earns $10 million a year.
LFM2.5-350MFreeLiquid AI's 350M on-device model for extraction and structured output.
Tiny AyaFreeCohere Labs' 3.35B open models for 70 languages, with GGUF files.
Gemma 4Open SourceChoose it when you need audio input on a small model: its E2B and E4B take audio natively, though their context is 128K against Qwen3.5 Small's 262K.
Questions about Qwen3.5 Small
What is Qwen3.5 Small?
Qwen3.5 Small is the Qwen team's set of four open models, from 0.8B to 9B parameters, that read text, images and video and run on your own hardware under Apache 2.0.
Who is Qwen3.5 Small for?
Qwen3.5 Small is best for developers who want a free, private model on their own hardware for reading documents and images, extraction, light agents with tool calls, or fine-tuning.
Is Qwen3.5 Small free?
Yes. Qwen3.5 Small has a free Open weights plan: 0.8B, 2B, 4B and 9B, plus Base versions.
Is Qwen3.5 Small open source?
Yes. Qwen3.5 Small's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Qwen3.5 Small alternatives?
The closest alternatives on Recaply are Falcon H1R 7B, LFM3, LFM 2.5, LFM2.5-350M, Tiny Aya and Gemma 4.
Where can I try Qwen3.5 Small?
On its official website, huggingface.co.
How much memory do the Qwen3.5 Small models need?
The BF16 weights alone take 1.75 GB for the 0.8B, 4.55 GB for the 2B, 9.3 GB for the 4B and 19.3 GB for the 9B, and long contexts need more on top. Qwen suggests a shorter context if you run out of memory.
How do I turn thinking on or off in Qwen3.5 Small?
The 4B and 9B think by default and the 0.8B and 2B answer directly. Set enable_thinking to false or true in the chat template options of vLLM or SGLang to switch; Qwen lists different sampling settings for each mode.
Can Qwen3.5 Small understand audio?
No. The small Qwen3.5 models read text, images and video frames and write text only. For speech input and spoken replies, Qwen's option is Qwen3.5-Omni, which runs only as an API on Alibaba Cloud.
Can I use Qwen3.5 Small in a commercial product?
Yes. All four sizes and their Base versions use the Apache 2.0 licence, which allows commercial use, changes and redistribution with no revenue limit, as long as you keep the licence and copyright notices.
Our verdict
Qwen3.5 Small is the open family to try first for local vision, extraction and light agent work on modest hardware.
Qwen's tables put the 9B ahead of much larger open models, Apache 2.0 allows commercial use, and users confirm its tool calling and OCR are strong for its size.
- Skip it for
- Serious coding or long agent runs, where users see invented details and runaway thinking; a 27B model does far better if your GPU can hold one.
- Check first
- Set Qwen's recommended sampling settings, or turn thinking off, before you judge it.
Qwen3.7-Plus
Qwen3.8-Flash
Qwen3.8-Max
Qwen3.6-Plus
Nemotron 3 Nano Omni
Mistral Small 4