The ultra-compact AI models of the Qwen3.5 series, 'Qwen3.5-0.8B', 'Qwen3.5-2B', 'Qwen3.5-4B', and 'Qwen3.5-9B', are all released at once.

Alibaba Cloud's Qwen team (Tongyi Lab) has released four small models in the ' Qwen3.5 Small ' series. The model sizes (number of parameters) are 0.8B (800 million), 2B (2 billion), 4B (4 billion), and 9B (9 billion), and each model is unique in that it can run on smartphones.

???? Introducing the Qwen 3.5 Small Model Series
Qwen3.5-0.8B · Qwen3.5-2B · Qwen3.5-4B · Qwen3.5-9B

✨ More intelligence, less computation.
These small models are built on the same Qwen3.5 foundation — native multimodal, improved architecture, scaled RL:
• 0.8B / 2B → tiny, fast,…

pic.twitter.com/90JfOM9k4T

— Qwen (@Alibaba_Qwen) March 2, 2026

Qwen3.5 is a large-scale language model developed by the Qwen team, and is characterized by its native support for a wide variety of languages and dialects, including Japanese.

Chinese AI 'Qwen3.5-397B-A17B' equivalent to GPT-5.2 appears as an open model - GIGAZINE

The newly announced Qwen3.5 Small series uses the same native multimodal model as Qwen3.5, and like Qwen3.5, it is built using an architecture that combines a Gated Delta Network and a Mixture of Experts (MoE) mechanism, as well as extended reinforcement learning. According to the Qwen team, 0.8B and 2B are 'small, fast, and ideal for edge devices,' 4B is 'a surprisingly powerful multimodal base for lightweight agents,' and 9B is 'compact, yet already closing the gap with much larger models.'

Below are the results of benchmark tests comparing Qwen3.5-9B (purple) and Qwen3.5-4B (light purple) with Qwen3-Next-80B-A3B-Thinking (brown), Qwen3-30B-A3B-2507 (blue), Qwen3-VL-30B-A3B (red), gpt-oss-120b (yellow), gpt-oss-20b (orange), GPT-5 nano (green), and Gemini 2.5 Flash-Lite (gray).

This table shows the benchmark scores for gpt-oss-120b, gpt-oss-20b, Qwen3-30B-A3B-Thinking, Qwen3-30B-A3B-Thinking-2507, Qwen3.5-9B, and Qwen3.5-4B. The Qwen3.5-9B and Qwen3.5-4B perform worse than the other models in Reasoning & Coding, but otherwise perform roughly equivalently.

pic.twitter.com/YucbP5ImfF

— Qwen (@Alibaba_Qwen) March 2, 2026

Looking at the scores for Text Recognition and Document Understanding and Video Understanding, Qwen3.5-9B and Qwen3.5-4B far exceed gpt-oss-120b and gpt-oss-20b.

pic.twitter.com/69X7Mc2pA5

— Qwen (@Alibaba_Qwen) March 2, 2026

The Qwen3.5 Small series has four models and each base model available on Hugging Face.

Qwen (Qwen)
https://huggingface.co/Qwen/models

The Qwen3.5 Small series can also be used with the app ' Ollama ,' which can run various AI functions locally.

qwen3.5:9b
https://ollama.com/library/qwen3.5:9b