Alibaba (Qwen)

Alibaba, through its cloud division Alibaba Cloud (Aliyun), has developed the Qwen family of large language models that has become one of the most significant open-source AI model families in the world. Qwen models span text, vision, code, and audio modalities, and have been adopted as the base for fine-tuning and deployment across thousands of applications globally. Combined with DeepSeek, Qwen represents the leading edge of China's open-source AI ecosystem and a challenge to the assumption that frontier AI capabilities will remain concentrated in Western labs.

Open-Source Scale

Qwen's open-weight releases have made it one of the most downloaded model families on Hugging Face. Qwen 2.5 and its successors have demonstrated competitive performance with frontier models on reasoning, coding, and multimodal tasks. Alibaba has released Qwen models optimized for agentic workflows, including tool use, code generation, and function calling — the capabilities that matter most for agentic engineering applications. This positions Qwen as a foundational layer for the agentic web in markets where Alibaba Cloud has infrastructure presence.

As of 6 October 2026 the Qwen organisation page on Hugging Face listed 468 model repositories and roughly 109,000 followers, and the open releases run from sub-1B models to a 2.4-trillion-parameter mixture-of-experts model. Older small models still dominate usage: on the organisation's download-sorted listing the top entry was Qwen3-0.6B (30.8M on the Hub's download counter), ahead of Qwen3-VL-8B-Instruct, Qwen3-8B and Qwen3.5-9B, with Qwen3.8-27B already at 6.8M.

Model Families as of October 2026

The following was checked on 6 October 2026 against the Qwen3.8 GitHub repository and the individual Hugging Face model cards; release dates are the ones those pages state.

Series	Open-weight models	Released	Licence
Qwen3.8	Qwen3.8-27B (dense, text/image/video); Qwen3.8-2.4T-A95B (MoE, 95B active, text-only)	14 and 12 August 2026	Apache-2.0 (27B); custom Qwen3.8-Max License (2.4T)
Qwen3.6	Qwen3.6-35B-A3B (MoE, 3B active); Qwen3.6-27B	16 and 22 April 2026	Apache-2.0 (35B-A3B card)
Qwen3.5	397B-A17B, 122B-A10B, 35B-A3B, 27B, 9B, 4B, 2B, 0.8B	February 2026 (per the Qwen3.5-9B card)	Apache-2.0 (9B card)

Alongside the main line, the organisation's GitHub pins Qwen3-VL (vision-language), Qwen3-Omni (end-to-end text, audio, image and video with speech output) and Qwen-Image-2.1, and the Hub carries Qwen3-Embedding and Qwen3Guard-Stream safety classifiers. Qwen Code, an open-source terminal coding agent, is the organisation's most-starred repository. Hosted versions are sold through Alibaba Cloud Model Studio, whose model list (page dated 28 September 2026) names qwen3.8-max, qwen3.7-plus and qwen3.8-flash; the weights for the Plus and Flash tiers were not found on the Hub.

Licensing: Apache-2.0 With One Large Exception

The models most people fine-tune — Qwen3.8-27B, Qwen3.6-35B-A3B and the Qwen3.5 sizes checked — are Apache-2.0. The flagship is different. Qwen3.8-2.4T-A95B ships under a bespoke Qwen3.8-Max License that permits commercial use but requires the model name to be displayed prominently in products with more than 100 million monthly active users or US$20 million in monthly revenue, and requires a separate licence for model-as-a-service providers whose aggregate revenue exceeds US$50 million over twelve months. For most adopters those thresholds never bind, but it means "Qwen is Apache-2.0" is no longer true of the whole family, and that the largest Qwen is an open-weight model in the narrow sense.

Architecture Changes

Since Qwen3.5 the family is no longer a standard transformer stack. The Qwen3.8-27B card describes 64 layers arranged as sixteen repetitions of three Gated DeltaNet (linear-attention) blocks followed by one Gated Attention block; Qwen3.6-35B-A3B uses the same 3:1 pattern over 40 layers with 256 experts, of which eight routed and one shared are active per token. Context is 262,144 tokens natively, with the cards describing extension to about one million. The models are vision-language by default (the 2.4T model is the text-only exception), run in thinking mode by default, and expose reasoning-effort levels; Qwen3.8-27B also preserves thinking across turns by default. Benchmark figures on the cards — for example 73.0 on Terminal Bench 2.1 and 61.7 on SWE-bench Pro for Qwen3.8-27B — are vendor-reported; no independent replication was located.

Fine-Tuning and Serving Support

The Qwen3.8 repository lists SGLang, vLLM, llama.cpp, MLX and TokenSpeed for inference, and Unsloth, Swift and LLaMA-Factory for training, covering supervised fine-tuning, DPO and GRPO. Unsloth's Qwen3.8 guide (opened October 2026) states that QLoRA on the 27B model fits in 24 GB of VRAM, that LoRA needs more than 36 GB and full fine-tuning about four times more, that Qwen3.8 loads under the qwen3_5 architecture name, and that at least 75% of the training mix should be reasoning-style data if reasoning ability is to be preserved. New architectures need matching runtime support: the Qwen3.6 card asks for vLLM 0.19.0 or SGLang 0.5.10 at minimum, and the Qwen3.5 card originally required builds from the main branch. A mid-sized, permissively licensed model with first-class tooling is what makes Qwen a common base for parameter-efficient fine-tuning and reinforcement fine-tuning work.

Alibaba Cloud and AI Infrastructure

Beyond models, Alibaba Cloud provides the compute infrastructure layer for AI deployment across Asia-Pacific, including custom AI accelerator hardware and inference optimization services. The integration of Qwen into Alibaba's commerce ecosystem — spanning Taobao, Tmall, AliExpress, and Alibaba.com — represents one of the largest real-world deployments of AI-powered digital commerce, with agents handling customer service, product recommendations, and logistics optimization at massive scale.

The Multipolar AI Landscape

Together with DeepSeek and other Chinese AI labs, Alibaba's Qwen models ensure that the global AI landscape remains multipolar rather than concentrated in a handful of Western companies. This has direct implications for the agentic web: as agents mediate more economic activity, the diversity of foundation models powering those agents affects competition, resilience, and the distribution of value across the global economy.