Alibaba Qwen (Tongyi)

Largest open-weights ecosystem (460+ open models, 3B downloads) plus a proprietary Max flagship; scale to 5-10T parameters, push recursive self-improvement, and tie models to Alibaba chips and cloud. Now a hybrid open/closed strategy with revenue-sharing licenses.

Founded: April 2023
Headquarters: Hangzhou, China
Now: Tech lead Junyang Lin quit 2026-03-03; the Qwen unit sat leaderless six months until Dayiheng Liu was named head (2026-09-22). Qwen3.8-Max ranks first on Artificial Analysis's Agentic Index but only seventh on its Intelligence Index (36Kr).

Key people

Dayiheng Liu (head of Qwen since 2026-09-22)
Eddie Wu (Alibaba CEO)
Zhou Jingren (Alibaba Cloud CTO, interim lead)
Junyang Lin (tech lead until 2026-03; founded Pragmatik Labs)

Latest launches

Qwen3.8-Max recursive self-improvement demos (Apsara 2026) Alibaba (Qwen) · 22 September 2026
Alibaba says Qwen3.8-Max ran 33 automated self-improvement cycles in a month (Artificial Analysis score 40 to 45) and cut a chip module's area 42%.
Qwen 4 (announced as in training, NOT released) Alibaba (Qwen) · 22 September 2026
At Apsara 2026 Alibaba said Qwen 4 is in training with no date; reports cite four tiers and 5-10T-parameter goals for Qwen 4.5 and 5.
Qwen-Image-2.1 (7B) Alibaba (Qwen) · 20 September 2026
Open image model with a 7B generation component that unifies text-to-image and editing, with native transparent-image generation and editing.
Qwen3.8-LiveTranslate (and Qwen3.5-LiveTranslate) Alibaba (Qwen) · 18 September 2026
Simultaneous interpretation model using an interleave architecture; average latency cut from 2.8 s to 2.3 s, with speaker naming.
Qwen3.8-Omni-Flash Alibaba (Qwen) · 17 September 2026
Natively multimodal agent model on the Qwen3.8-Next MoE architecture with a 1M-token context, built to plan and finish tasks, not just perceive.
Qwen-Drive-1.0 (4B) Alibaba (Qwen) · 2 September 2026
Vision-language foundation model for autonomous driving, built on Qwen3.5-4B, unifying 3D perception with visual question answering.
Qwen3.8-Max-0902 Alibaba (Qwen) · 2 September 2026
September snapshot of the hosted Qwen3.8-Max with improved coding and multi-agent collaboration.
Qwen3.8-Flash-Next Alibaba (Qwen) · 26 August 2026
125B-parameter (6B active) multimodal MoE that previews the Qwen4 architecture, built on Gated DeltaNet plus Qwen Sparse Attention, with 51B of N-gram embeddings.

Launches by year

106 launches and papers since 3 August 2023, oldest first within each year.

2026 42: Qwen3-VL-Embedding and Qwen3-VL-Reranker (2B, 8B), Qwen3-TTS (0.6B, 1.7B) with 12Hz tokenizer, Qwen3-Max-Thinking, Qwen3-ASR (0.6B, 1.7B) and Qwen3-ForcedAligner, Qwen3-Coder-Next (80B-A3B), Qwen-Image-2.0, Qwen3.5-397B-A17B and Qwen3.5-Plus, WebWorld (8B, 14B, 32B), Qwen3.5 medium and small models (122B-A10B to 0.8B), Qwen3.5-Omni (Plus, Flash and Realtime), Qwen3.6-Plus, Wan2.7 (T2V, I2V, R2V, video edit), HappyHorse-1.0, Qwen3.6-Max-Preview, Qwen3.6-35B-A3B and Qwen3.6-27B, FlashQLA (linear-attention kernels), Qwen-Scope (sparse autoencoders for Qwen3 and Qwen3.5), Qwen3.7-Max, Qwen-VLA (vision-language-action generalist), Qwen3.7-Plus, Qwen-Robot Suite (RobotNav, RobotManip, RobotWorld), Qwen-AgentWorld (language world models for agents), Wan-Dancer-14B, Qwen-Audio-3.0 and 3.1 (TTS, ASR, realtime duplex), Qwen-Image-3.0, Qwen3.7-Flash, Qwen-MM-Plugins (multimodal plugins for agent harnesses), QwenWork (enterprise agent workspace), Qwen3.8-Max, Wan-Animate-2 (and Lite), Qwen3.8-2.4T-A95B (open weights), Wan 3.0 (and Wan 3.0 Prime), Qwen3.8-27B, Wan3.0 (video), Qwen3.8-Flash-Next, Qwen3.8-Max-0902, Qwen-Drive-1.0 (4B), Qwen3.8-Omni-Flash, Qwen3.8-LiveTranslate (and Qwen3.5-LiveTranslate), Qwen-Image-2.1 (7B), Qwen 4 (announced as in training, NOT released), Qwen3.8-Max recursive self-improvement demos (Apsara 2026)
2025 46: Qwen2.5-Math-PRM and ProcessBench, Global-batch load balancing for MoE training, Qwen2.5-VL (3B, 7B, 32B, 72B), Qwen2.5-1M (open models with 1M-token context), Qwen2.5-Max, Wan 2.1, Wan2.1 (T2V 1.3B and 14B, I2V 14B), QwQ-Max-Preview, QwQ-32B, Qwen2.5-Omni (3B, 7B), QVQ-Max-Preview, Qwen3 (0.6B to 235B-A22B), Wan2.1-VACE (1.3B, 14B), Qwen3 Technical Report, ParScale: Parallel Scaling Law for Language Models, Qwen3 Embedding and Reranker, Qwen VLo, Qwen Code (open-source coding CLI), Qwen3-235B-A22B-Instruct-2507 and Thinking-2507, Qwen3-Coder-480B-A35B and Qwen Code, Qwen-MT, GSPO: Group Sequence Policy Optimization, Wan 2.2, Wan2.2 (T2V-A14B, I2V-A14B, TI2V-5B), Qwen3-Coder-30B-A3B-Instruct (Coder-Flash), Qwen-Image, Qwen-Image (20B MMDiT), Qwen-Image-Edit, Wan2.2-S2V-14B, Qwen3-Max-Preview, Qwen3-ASR-Flash, Qwen3-TTS-Flash and Qwen3-LiveTranslate-Flash (APIs), Qwen3-Next-80B-A3B (Instruct and Thinking), Wan2.2-Animate-14B, Qwen3-Omni-30B-A3B, Qwen-Image-Edit-2509, Qwen3Guard (Gen and Stream; 0.6B, 4B, 8B), Qwen3-VL (235B-A22B first; 2B to 32B and 30B-A3B in October), Wan2.5-Preview, Qwen3-Max, Qwen app (public beta), Z-Image, Wan2.6 series (T2V, I2V, R2V, T2I, image), Wan 2.6, Qwen-Image-Layered, Qwen-Image-Edit-2511, Qwen-Image-2512
2024 14: Qwen-VL-Plus and Qwen-VL-Max (API), Qwen1.5 (0.5B to 72B; later 32B and 110B), Qwen1.5-MoE-A2.7B, CodeQwen1.5-7B, Qwen-Max-0428, Qwen2 (0.5B, 1.5B, 7B, 57B-A14B MoE, 72B), Qwen2-Math and Qwen2-Audio, Qwen2-VL (2B, 7B, 72B), Qwen2.5 (0.5B-72B) with Qwen2.5-Coder and Qwen2.5-Math, Qwen2.5-Coder series (up to 32B), Qwen2.5-Turbo (1M-token context), QwQ-32B-Preview, Qwen2.5 Technical Report, QVQ-72B-Preview
2023 4: Qwen-7B and Qwen-7B-Chat, Qwen-VL and Qwen-VL-Chat, Qwen-14B and the Qwen Technical Report, Qwen-72B, Qwen-1.8B and Qwen-Audio

People moves

Of the moves the atlas tracks, 0 joined and 2 left.

Lin Junyang 12 August 2026
Alibaba (Qwen) to Pragmatik Labs
Founder, Pragmatik Labs; ex-Qwen technical lead
Lin Junyang 4 March 2026
Alibaba (Qwen) to unknown
Technical lead, Qwen

Sources

This profile was checked against its sources on 6 October 2026. How we check