Alibaba
Operates Qwen models and Bailian (阿里云百炼). Qwen3.8-Max is the strongest open-weights Chinese model (2.4T params, $2/$6 per 1M).
Operates Qwen models and Bailian (阿里云百炼). Qwen3.8-Max is the strongest open-weights Chinese model (2.4T params, $2/$6 per 1M).
All Alibaba Plans(7)
Models from Alibaba(1)
Alibaba History Timeline(30)
Purported Qwen 4 spotted in Arena anonymous testing, unconfirmed by vendor
The anonymous model Kalmia was inferred by community analysis to correspond to Qwen 4; Alibaba has not confirmed identity or timing.
SourceChina weighs 23 AI tech export controls
Oct 16 China Ministry of Commerce weighs 23 AI tech export controls including Qwen open-source model weights
SourceAlibaba Qwen launches Personal Agent Qwen Go!
Alibaba's Qwen on October 13 opened the Qwen Go! Personal Agent experience request page in the Qwen app, capable of autonomously handling flight delay compensation, Taobao price protection refunds, auto maintenance booking, sports venue reservation, and cross-app price comparison.
SourceQwen-Image-2.1-Turbo open-weight release
Open-weight 7B Qwen-Image-2.1-Turbo released on October 9, with denoising steps cut from 40 to 8 and hosted API at CNY 0.10 per image.
SourceAlibaba ships Qwen-Audio 3.1 with up to 50% price cut
Qwen-Audio 3.1 improves recognition accuracy, low-SNR performance, and speaker diarization; Qwen-Audio voice API pricing cut by up to 50% across the line.
SourceQwen Book + AI Glasses N1 + AI clip-on earbuds trio launches
9/24 Hangzhou Cloud Summit launch, sales begin 10/13; Qwen Book built on OS as Harness with Omarchy open-source collaboration
SourcePingtouge Zhenwu V900 AI chip launched, single cluster scale to 500K cards
9/24 3x compute of M890, 215GB memory, 1.2TB/s interconnect; Qwen4 already in training
SourceZhenwu V900 chip and Qwen 5-10 trillion-parameter roadmap unveiled
At Apsara Conference on September 23 Alibaba unveiled T-Head's Zhenwu V900 (3x M890 performance, 216GB memory, 1.2TB/s interconnect, 500,000-card supercluster) and a Qwen 5-10T-parameter roadmap. Mass production starts Q1 2027; Alibaba Cloud targets 20GW global capacity by 2032.
SourceAlibaba Qwen-Audio 3.1 family launches with up to 95% ASR price cut
At the Apsara Conference on September 23, Alibaba Qwen launched the Qwen-Audio 3.1 family with five models including ASR-Next multi-speaker recognition and TTS-Next diffusion synthesis. The Beijing node Qwen-Audio-3.1-ASR-Flash works out to roughly $0.016 per audio hour, about one-ninth of Meta Muse Voice Transcribe.
SourceAlibaba Qwen launches Qwen Intelligence mobile agent suite
Alibaba Qwen on September 22 launched Qwen Intelligence at the Apsara Conference, including Mobile Planner, Mobile-Use (90% end-to-end success), and Mobile Creative agents, alongside open benchmarks MobilePA-Bench and others. Honor Magic9 on September 28 will be the first phone to officially carry the solution.
SourceQwen team open-sources Qwen-Image-2.1 unified image model
Alibaba's Qwen team open-sourced Qwen-Image-2.1 on September 20, a 7B / 32-layer Single-Stream DiT that unifies text-to-image, image editing, transparent RGBA, and up to 10 reference images; Day-0 support covers Diffusers, ComfyUI, vLLM-Omni, and SGLang.
SourceAlibaba Qwen3.8-27B becomes most-liked open model on Hugging Face
On September 16, Alibaba's Qwen3.8-27B became the most-liked open-source model in Hugging Face history.
SourceAlibaba open-code-review tops GitHub Trending
Deterministic pipeline + LLM agent hybrid, ~1/9 token use, 31,689 stars.
SourceQwen3.8 early-access preview
Sept 12 release, Dynamic Router v4 + Sparse-Tiled Attention + KV cache quantization; 128K+ context memory cut 40%
SourceQwen AI glasses first revealed
Sept 10 Bund Summit AI payments exhibit, possible iris-recognition payments
SourceAccio AI Agent e-commerce task cost
107 tasks $3.69 average, 50%+ cheaper than Codex/Claude Code
SourceQwen3.8-Max-0902 tops Code Arena WebDev
Sept 2 post-training refresh, 2.4T-parameter MoE, 1M context, 1691 points
SourceAlibaba open-sources Qwen3.8-Flash model with 90% lower training cost
Qwen3.8-Flash uses next-gen architecture, 125B params activate 6B, 1 yuan input / 3 yuan output per million tokens, prototype for Qwen4
SourceAlibaba QwenWork international public beta opens
QwenWork international beta via qwenwork.ai, integrated with Slack/Notion, ranked first in Jefferies evaluation
SourceThomson Reuters launches Thomson-1 on Qwen3.5
For table analysis and document review, gradually reducing reliance on Claude. Stanford LegalBench score 0.823.
SourceQwen App and PC support Alibaba Cloud Token Plan integration
Users can bind an API Key to consume Token Plan subscription in Work Assistant for complex tasks.
SourceQwen3.8-27B receives AMD Day 0 support
Runnable on AMD PCs/workstations with 24GB+ VRAM. Artificial Analysis Intelligence Index score 52.
SourceJack Ma and management buy HK$800M in Alibaba shares
Ma personally bought over HK$600M; Tsai and Wu combined HK$200M. In response to the HK$80B placement.
SourceAlibaba prices HK$80B Hong Kong placement, 100% to AI buildout
710M new shares at HK$112.70 to non-US persons outside the US, closing expected 8/26. Use of proceeds covers AI compute infrastructure expansion, large-model R&D iteration and commercial rollout. Alibaba's first follow-on placement since its 2019 Hong Kong dual primary listing.
SourceHK$80B share placement, 100% to full-stack AI
HK$80B share placement, 100% to full-stack AI
SourceAlibaba Tongyi App completes brand upgrade and becomes Qianwen at v5.0.0
Tongyi Qianwen, Qianwen 2C, Qianwen Office and other brand lines were consolidated under the unified Qianwen brand. On Aug 20 Alibaba merged Qianwen and AI Labs into an AI Labs and Applications division, upgrading AI from technology middle-office to a standalone business line.
SourceAlibaba open-sources Qwen-UI-Agent GUI agent foundation model
Qwen-UI-Agent covers mobile, desktop, web and deep search; trained on 100+ real devices and 150+ apps; MobileWorld 82.1, MobileWorld-Real 92.2, AndroidDaily 97.5, WebArena 73.6 (rank 1), OSWorld Verified 79.5. Includes 3D director mode, online RL on 100+ step trajectories. Technical report and code public.
SourceQwen Office integrates WeCom, covering all three major Chinese collaboration platforms
Aug 19 Qwen Office connected WeCom, DingTalk, Feishu, plus overseas Slack
Source