has done it with promising pocket AI models

The latest models from OpenAI, Anthropic or Google are fantastic, no doubt, but they have a problem: they are gigantic, so the only way to use them is to use the chatbots of these companies. But while those companies focus on that approach, Alibaba just surprised us with something fascinating.

The allure of tiny AI models. This Chinese technology giant just launched the family of “Qwen 3.5 Small Models”, which is made up of four variants of open models with really small sizes. Thus, we have a “dwarf” model with 800 million parameters (0.8B), another with 2,000 million (2B), a third with 4,000 million (4B) and the last one with 9,000 million (9B). There are no official figures for the number of parameters for GPT 5.3, Opus 4.6 or Gemini 3.1, but it is very likely that they are all around 500B or far higher.

Tiny but bully. The first two models are designed for prototyping and deployment in very modest devices and in which battery life is a priority, because their consumption is also very tight. Meanwhile, Qwen3.5-4B is a multimodal model for lightweight AI agents that supports a context window of up to 262,144 tokens. The latter, for example, has a size of less than 3 GB in its 4-bit quantized version, which makes it usable even on mobile phones. Even more interesting is the “eldest” of the family.

The best essences… The latest of these models, Qwen3.5-9B, is really promising. It is a reasoning model that according to its creators surpasses nothing less than gpt-oss-120B, the open OpenAI model that is 13.5 times larger and that until now was a great reference in this field. All of these models are open weights, and can be found in both Hugging Face as in ModelScope in its different variants.

A new approach. In these models Alibaba has made some changes and makes use of what they call Efficient Hybrid Architecture in which they combine a new type of attention algorithms (Gated Delta Networks) with the already known Mixture-of-Experts (MoE). This approach allows you to avoid the “memory wall” problem that affects small models.

Promising returns. The results of the benchmarks published by Alibaba are really striking. Both Qwen3.5-4B and Qwen3.5-9B make a notable leap in efficiency, especially in multimodal tests—these models are capable of using images as input—and reasoning tasks. Thus, in the MMMU-Pro visual reasoning test, Qwen3.5-9B left Gemini 2.5 Flash lite behind, and in the GPQA reasoning test the Alibaba model 9B even managed to leave gpt-oss-120b behind.

Alibaba surpasses itself. Paul Couvert, popularizer of AI, showed his enthusiasm in Xwhere he explained that at least according to these benchmarks Qwen3.5-4B was as powerful as Qwen3-Next-80B-A3B-Thinking, which until not long ago was considered a marvel but which had a notable size.

Models for your laptop and your mobile. These models are especially striking because they give the option for practically anyone to use them on their laptop or mobile phone (or integrated into a browser!). In all cases the advantages are clear: you do not depend on the cloud, so you can use them offline, and our conversations do not go through any server, so “everything stays at home” and when using these models, the chats are private.

Only Google seems to follow suit. Of the Western AI majors, only Google seems to be interested in small models. Gemma 3 270M was a surprising version launched in August 2025. Microsoft also has its Phi-4 December 2024, but beyond that there are few examples. OpenAI launched gpt-oss-20B and gpt-oss-120B in August 2025 and showed some interest in this type of scenario, but there has been no news since then. There are startups like Liquid that have a eye-catching LFM2.5 with a variant of only 1.2B, but here Alibaba seems unstoppable with its commitment to small. At least, for now.

In Xataka | If the question is which of the big tech companies is winning the AI ​​race, the answer is: none