Latest23 Sep 2026Audio/voice model
Qwen Audio 3.1
A five-model voice family: speech recognition, text-to-speech, a real-time duplex voice model, and 'Next' versions that also understand and generate non-speech sound. It launched with large voice-API price cuts, which press put at up to 95% for recognition, about 85% for real-time and about 70% for speech synthesis.
- Models: ASR, TTS, Realtime, ASR-Next, TTS-Next
- ASR: 30 languages and 16 Chinese dialects (reported)
- Realtime: full-duplex talk with interruptions and tool calls
- Realtime-Plus: 262,144-token context, eight system voices (Model Studio)
AccessAPI on Alibaba Cloud Model Studio and the Qwen API platform (ASR-Next API not yet open per press)
20 Sep 2026Image model
Qwen-Image-2.1
Alibaba's newest open image generator and editor. It adds native transparent (RGBA) output, editing with up to ten reference images, better text rendering and portrait lighting, and had same-day support in Diffusers, ComfyUI, vLLM-Omni and SGLang. A hosted Pro version followed on Alibaba Cloud on 2 October 2026.
- 7B visual generation component (32 single-stream DiT layers)
- Text-to-image and multi-reference editing
- Native transparency (RGBA)
- Up to 2752x1536 output
AccessOpen weights on Hugging Face and ModelScope (Qwen Research License); hosted qwen-image-2.1-pro on Model Studio from 2026-10-02
Aug 2026Open-weights LLMmonth only
Qwen3.8-27B and Qwen3.8-Flash
The smaller members of the Qwen3.8 generation. Qwen3.8-27B is a dense vision-language model for coding and office work; Qwen3.8-Flash-Next is an open preview of the architecture planned for Qwen4, and Qwen3.8-Flash is its production API version with 1M context by default.
- 27B: dense, image and video input, 262,144 native context (to 1M)
- Flash-Next: 125B total / 6B active plus 51B n-gram embeddings
- Flash-Next: SWE-bench Pro 62.5 (lab model card)
- Thinking on by default, can be turned off (27B)
AccessQwen3.8-27B: open weights, Apache-2.0; Qwen3.8-Flash-Next: open weights, Qwen Community License 1.0; qwen3.8-27b and qwen3.8-flash via Model Studio API
Aug 2026Video modelmonth only
Wan3.0
Alibaba's third-generation Wan video generator. It makes clips up to 30 seconds, twice its predecessor, and can take text, images, video, audio and even office documents as input in one prompt. Unlike Wan 2.x, it has not been released as open weights.
- Clips up to 30 seconds
- Up to 1080p
- Reference, editing, replication and driving modes (Model Studio)
- Up to 10 images, 5 videos and 5 audio clips per prompt (reported)
Accesswan.video and Alibaba Cloud Model Studio API; no downloadable weights (per press)
3 Aug 2026Frontier LLM
Qwen3.8-Max
Alibaba's largest model and current flagship, and the first Max-class Qwen model released as open weights. Alibaba says it ran a real software project on its own for 16 days. The open checkpoint is text-only with thinking always on; a September API snapshot added native vision.
- 2.4T total / 95B active MoE, hybrid attention
- 262,144 native context, extendable to about 1M
- Open checkpoint: text-only, thinking always on
- qwen3.8-max-0902 snapshot adds native vision (Model Studio)
AccessAPI on Alibaba Cloud Model Studio; QwenWork; open weights as Qwen3.8-2.4T-A95B on Hugging Face (Model Studio lists it 2026-08-12) under the Qwen3.8-Max License
Price$2 input / $6 output per 1M tokens (0-1M tier, Model Studio international)
21 May 2026Frontier LLM
Qwen3.7-Max
The Qwen3.7 generation's top model, pitched by Alibaba as a flagship for the agent era of coding, debugging and office automation. Alibaba says it can run more than 1,000 tool calls in one task and work on its own for up to 35 hours. A June snapshot added visual understanding.
- 1,000+ tool calls in one task (Alibaba claim)
- Up to 35 hours autonomous execution (Alibaba claim)
- Visual understanding added in 2026-06-08 snapshot
- Thinking mode
AccessAPI on Alibaba Cloud Model Studio (qwen3.7-max-preview from 2026-05-20); no open weights
Price$2.5 input / $7.5 output per 1M tokens (0-1M tier, Model Studio international)
2 Apr 2026Frontier LLM
Qwen3.6 family (Plus, Max-Preview, 35B-A3B, 27B)
Qwen3.6 focused on agentic coding: the hosted Plus model plans, tests and fixes code itself and turns screenshots or sketches into front-end code. Alibaba later opened two smaller models, a 35B mixture-of-experts model with 3B active and a 27B dense model.
- Plus: 1M-token context by default
- 35B-A3B: 35B total / 3B active, text, image and video input
- 35B-A3B: SWE-bench Verified 73.4 (lab model card)
- Works with Claude Code, Cline and OpenClaw (Alibaba)
AccessQwen3.6-Plus via Model Studio, Qwen Chat and Wukong; Qwen3.6-Max-Preview API; open weights for Qwen3.6-35B-A3B and Qwen3.6-27B, Apache-2.0
PriceQwen3.6-Plus: $0.5 input / $3 output per 1M tokens (up to 256K); $2 / $6 (256K-1M)
16 Feb 2026Open-weights LLM
Qwen3.5 family
A natively multimodal generation released as open weights in sizes from 0.8B up to a 397B mixture-of-experts flagship with 17B active. It mixes linear and full attention for faster inference, expanded language support to 201 languages and dialects, and is offered hosted as Qwen3.5-Plus and Qwen3.5-Flash.
- Flagship 397B total / 17B active
- Other sizes: 122B-A10B, 35B-A3B, 27B, 9B, 4B, 2B, 0.8B
- Text, image and video input
- 262,144 native context, extendable to about 1M
- 201 languages and dialects
AccessOpen weights on Hugging Face (Apache-2.0); Qwen3.5-Plus and Qwen3.5-Flash via Model Studio; Qwen Chat
PriceQwen3.5-Plus: $0.4 input / $2.4 output per 1M tokens (up to 256K)