Every significant Tongyi Qwen model, product and organisational update from the April 2023 announcement through August 2026 - newest first. Highlighted entries mark the current generation.
A dense 27B model with 262K native context (extensible to 1M via YaRN), multimodal input across text, images and video, and a thinking mode that is on by default but can be toggled per query. Qwen claims it outperforms the much larger hosted Qwen3.7-Plus on coding and office tasks, and it immediately became the default recommendation for local and single-GPU deployment.
vLLM shipped support on the day of release, making the largest open Qwen servable on standard inference stacks without waiting for a release cycle.
The first Max-class Qwen ever published with downloadable weights: 2.4T total parameters, 95B active, 92 layers, a 512-expert MoE activating 11 experts per token, and a hybrid of Gated DeltaNet linear attention with gated full attention. Text-only, always in thinking mode, 262,144 native context extensible to 1,010,000 tokens.
The hosted flagship launched with a 1M-token context window, native text/image/video input, a reasoning_effort control and built-in tools. Alibaba positioned it for multi-day autonomous coding, research reproduction and parallel agent orchestration, claiming leadership on agentic computer-use benchmarks. Priced at $2 in / $6 out per million tokens.
Speech recognition with improved context consistency, custom hotwords and domain-term handling; reported recall above 93% on medical and industrial terminology.
Hosted text-to-speech across 16 languages with expressive inline delivery controls and continuous output of up to three minutes.
Image generation handling complex layouts, precise text rendering in 12 languages, full LaTeX pages and more than 100 art styles.
Early access through Alibaba's Token Plan and partner platforms, two weeks ahead of general availability.
Hosted multimodal agent model with a one-million-token context window and substantially stronger visual understanding than the 3.5 line.
Proprietary flagship targeting coding, office automation, tool use and autonomous workflows. The 3.7 generation never received an open-weight release - it was skipped in favour of 3.8.
A dense 27B Apache 2.0 model that beat the previous generation's 397B-A17B flagship on coding suites - an early sign of how fast the small-dense line was improving.
Natively multimodal mixture of experts under Apache 2.0, with roughly 3B parameters active per token.
Proprietary model made available only through Alibaba's own chatbots and Alibaba Cloud.
The largest open model of the 3.5 generation arrived (88.4 GPQA Diamond, 86.7 Tau2-Bench). In the same month Qwen division head Lin Junyang departed, prompting public debate about the future of Alibaba's open-weight commitment.
A full ladder of open-weight models from 0.8B to 122B-A10B, all with 256K context under Apache 2.0, handling text, images, video, reasoning and agent tasks natively rather than through bolted-on adapters. The hosted Qwen3.5-Plus and Qwen3.5-Flash tiers launched alongside.
A mobile update connected Qwen to Taobao and Taobao Instant Commerce, Alipay, Fliggy and Amap, turning the assistant into an executor: ordering food, booking travel, placing phone calls with transcripts, and batch-processing up to 100 documents. Alibaba also reorganised its AI operations into a group called Token Hub under CEO Eddie Wu.
Alibaba's consumer assistant entered public beta in China and passed 100 million monthly active users within two months.
Qwen's first trillion-parameter model, hosted only, launched as an Instruct release with a Thinking variant to follow.
Vision-language models with stronger visual perception plus spatial and video reasoning, in separate Instruct and Thinking variants.
End-to-end multimodal model accepting text, images, audio and video and streaming both text and speech back.
A 20B open-weight text-to-image and editing model with unusual attention to rendering readable text - in both Chinese and English - inside generated images.
Agentic coding models at 480B-A35B and 30B-A3B, paired with an open-source command-line tool for repository-scale, multi-step development work.
Dense models at 0.6B, 1.7B, 4B, 8B, 14B and 32B plus 30B-A3B and 235B-A22B MoE variants, all Apache 2.0, trained on 36 trillion tokens across 119 languages. This is the generation that brought hybrid thinking / non-thinking behaviour into the core family.
The first open Omni model: multiple input modalities in, streamed text and speech out, enabling real-time voice conversation.
An open reasoning model trained with reinforcement learning for maths, coding and problem-solving.
A large-scale mixture-of-experts model, released days after DeepSeek's R1 shook the market.
Vision-language family at 3B, 7B, 32B and 72B with much better document parsing, chart reading, long video understanding and localisation.
The experimental model that opened Qwen's dedicated reasoning branch and was the first open model widely benchmarked against OpenAI's o1.
The coding family expanded to six sizes from 0.5B to 32B, covering 92 programming languages.
A coordinated release of general, Coder and Math models with much better structured output - the generation that made Qwen a default open baseline worldwide.
Added dynamic image resolution, video understanding and visual localisation at 2B, 7B and 72B.
Apache 2.0 audio-language model interpreting speech, sound and music directly, without a separate transcription step.
Second generation, with better code, maths and multilingual ability and 128K context options.
The first dedicated Qwen coding family, separating code-focused training from general chat.
Wider size range and direct integration with Hugging Face Transformers; later added Qwen's first public mixture-of-experts model.
The first generation extended in both directions at once - a large flagship and a compact model for constrained hardware.
A mid-sized model plus an open-source agent framework for tool-using applications. Tongyi Qianwen also opened to the Chinese public this month after regulatory approval.
The first vision-language models, supporting image understanding, text reading and visual grounding.
The first publicly released Qwen open weights - base and chat variants. The beginning of the open-weight strategy that now accounts for over 40 million downloads.
Alibaba Cloud unveiled its large language model with enterprise beta access in China.
Qwen3.6-27B beat the previous generation's 397B flagship on coding; Qwen3.8-27B beats the hosted 3.7-Plus. If you deployed a huge open model six months ago, a dense 27B is probably now cheaper and better.
2026 opened with proprietary Plus and Max tiers and worried commentary about Alibaba retreating from open source. It ended the first half by publishing Max-class weights - under a bespoke licence rather than Apache 2.0.
Every 2026 release is framed around autonomy: computer use, terminal tasks, long-horizon execution, tool calling, and in the consumer app, actually placing orders and phone calls.