The full Tongyi Qwen lineup as of August 2026 — hosted flagships, open weights, multimodal and media models, and the earlier generations still in wide use. Filter by what you need.
2 Aug 2026
The current hosted flagship. A 2.4T-parameter MoE with ~95B active parameters, aimed at multi-day autonomous coding, research reproduction and parallel agent orchestration. Uses vision during execution to inspect its own intermediate output and self-correct.
12 Aug 2026
The first Max-class Qwen ever published with downloadable weights. Hybrid architecture: 92 layers, hidden size 8192, Gated DeltaNet linear attention combined with gated full attention, and a 512-expert MoE layer activating 11 experts per token (10 routed + 1 shared). Text-only, and thinking mode is always on.
14 Aug 2026
The new default local Qwen. A dense 27B model that Alibaba says beats the much larger hosted Qwen3.7-Plus on coding and office tasks, with stronger independent planning for agent work. Thinking mode is on by default and can be switched off per query.
19 Jul 2026
Early-access build of Qwen3.8-Max, distributed through Alibaba's Token Plan and partner platforms ahead of general availability.
1 Jun 2026
Hosted multimodal agent model with a one-million-token context window and notably stronger visual understanding than the 3.5 line. Positioned as the workhorse tier below Max.
20 May 2026
Proprietary flagship focused on coding, office automation, tool use and autonomous workflows. Never shipped open weights - the 3.7 generation was skipped for open release in favour of 3.8.
2026
The cheapest hosted tier, aimed at high-volume classification, extraction and routing workloads that still need a large context budget.
22 Apr 2026
Dense 27B model that, at release, beat the far larger Qwen3.5-397B-A17B on coding suites. Superseded by Qwen3.8-27B.
15 Apr 2026
Natively multimodal MoE with only ~3B parameters active per token - the cost/quality sweet spot of the 3.6 generation for self-hosting.
1 Apr 2026
Proprietary tier restricted to Alibaba's own chatbots and Alibaba Cloud customers.
Mar 2026
The largest open model of the 3.5 generation and, for several months, the strongest downloadable Qwen. Still a common choice for self-hosted frontier-ish workloads.
16 Feb 2026
The first natively multimodal Qwen generation, shipped as a full ladder: 0.8B and 2B for edge and mobile agents, 4B and 9B for local coding, 27B for single-GPU production, then 35B-A3B and 122B-A10B mixtures of experts. Architecture is roughly 75% Gated DeltaNet linear-attention layers and 25% full softmax attention.
16 Feb 2026
The hosted tiers of the 3.5 generation. Flash remains one of the cheapest hosted models anywhere with a genuine million-token context budget.
2026
Thinker-Talker MoE architecture of roughly 100B parameters, taking text, image, audio and video and returning streamed text and speech. Supports 113 languages.
22 Jul 2026
Image generation with an unusual strength: accurate typography inside the image. Handles complex layouts, precise text rendering in 12 languages, full LaTeX pages, and more than 100 art styles.
23 Jul 2026
Hosted text-to-speech in Flash and Plus tiers, with expressive inline controls over delivery. Supports 16 languages and continuous output of up to three minutes.
31 Jul 2026
Speech recognition tuned for real deployments: better context consistency, custom hotwords and domain vocabulary, with medical and industrial term recall reported above 93%.
4 Aug 2025
The 20B open-weight text-to-image and image-editing model that established Qwen's focus on commercial-grade Chinese and English text rendering inside generated images.
24 Sep 2025
The first trillion-parameter Qwen, hosted only. Launched as an Instruct model with a Thinking variant following.
22 Sep 2025
Vision-language line with stronger visual perception plus spatial and video reasoning, split into Instruct and Thinking variants. Qwen3-VL-2B-Instruct alone passed 18 million downloads.
21 Sep 2025
End-to-end multimodal model accepting text, image, audio and video, streaming both text and speech output for real-time conversation.
22 Jul 2025
Agent-oriented coding models in 480B-A35B and 30B-A3B sizes, shipped alongside Qwen Code, an open-source CLI for repository-scale, multi-step development tasks. The 30B-A3B build remains a favourite for Apple Silicon via MLX.
28 Apr 2025
The generation that brought hybrid thinking / non-thinking behaviour into the core family. Dense models at 0.6B, 1.7B, 4B, 8B, 14B and 32B plus 30B-A3B and 235B-A22B mixtures of experts, trained on 36 trillion tokens across 119 languages.
27 Mar 2025
The first open Omni model - multiple input modalities in, streamed text and speech out. The basis for real-time voice chat in the Qwen apps.
6 Mar 2025
Qwen's dedicated reasoning branch, trained with reinforcement learning for maths, code and problem-solving. Its November 2024 preview was the first open model widely compared to OpenAI's o1.
26 Jan 2025
Vision-language family at 3B, 7B, 32B and 72B with much better document parsing, chart reading, long-video understanding and localisation. Processes images at arbitrary resolution.
Sep–Nov 2024
The generation that made Qwen a default open baseline: a coordinated release of general, Coder and Math models with much better structured output. Qwen2.5-Coder shipped in six sizes from 0.5B to 32B and covers 92 programming languages.
2024
Qwen1.5 (Feb 2024) integrated directly with Hugging Face Transformers and introduced Qwen's first public MoE. Qwen2 (Jun 2024) improved code, maths and multilingual ability and introduced 128K context options.
2023
Qwen-7B and Qwen-7B-Chat (3 Aug 2023) were the first public open weights, followed by Qwen-VL (22 Aug), Qwen-14B and the Qwen-Agent framework (25 Sep), and Qwen-72B / Qwen-1.8B (30 Nov).
The generation. Higher is newer; generations ship roughly every one to three months in 2026.
Size. A single figure means a dense model.
X-AY means a mixture of experts with X total parameters and Y active per token —
cost tracks the active figure, quality tracks somewhere in between.
Hosted service tiers, not open weights. Max is the frontier tier, Plus the balanced workhorse, Flash the cheap high-volume option.
Specialisation: code, vision-language, end-to-end multimodal, speech and image generation respectively.
Post-training. Base is raw pretrained, Instruct follows instructions without visible reasoning, Thinking reasons before answering. Newer generations often fold both into one model with a toggle.
Early access ahead of general availability. Expect the production release to differ in price, quotas and sometimes quality.
| Licence | Typical models | Commercial use | Notes |
|---|---|---|---|
| Apache 2.0 | Most open Qwen3, 3.5, 3.6, 3.8-27B releases | Yes | The permissive default; modification and redistribution allowed with attribution. |
| Qwen3.8-Max licence | Qwen3.8-2.4T-A95B | Model-specific | A bespoke licence attached to the Max-class open release — read it before deploying. |
| Tongyi Qianwen licence | Several earlier large models | Usually, with conditions | Source-available rather than fully open; may include usage thresholds. |
| Qwen Research licence | Selected research checkpoints | No | Non-commercial research use only. |
| Proprietary | All Max / Plus / Flash hosted tiers | Via API terms | No weights; use is governed by Alibaba Cloud's service agreement. |