Over half of users are enterprises โ and an independent eval ranked it #1 among 8 mainstream agents. Here's what it costs to build your own on Alibaba Cloud.
๐ฅ 30M users, >50% enterprise โ Model Studio from $0.15/M input โ Cache hits $0.016/M ๐ Free credits for new usersOn October 14, 2026, Alibaba’s AI office product Qwen Office launched Qwen Office E-Commerce Edition, now in invite-only beta — merchants switch to “E-Commerce Ops” mode inside Qwen Office to try it. Positioned as an independent third-party e-commerce agent open to all platforms, the first release bundles four professional features: cross-platform operations dashboard, trend insight & selling-point analysis, bulk viral-content generation, and cross-platform ROI optimization, with integrations to BizSass (็ๆๅ่ฐ), Youzan, Qianniu, Alimama Wantai, FastMoss and LinkFox. Qwen Office says e-commerce users now make up ~20% of paying customers, running 38 complex tasks per week on average; one women’s-shoes merchant cut new-store setup from ~30 days to 7 and initial spend from ¥30,000–50,000 to just a few hundred yuan in token fees — a shift from heavy “make-then-sell” operations to asset-light “validate-then-make”. For teams building on Alibaba Cloud, the same Qwen3.8-Flash stack via Model Studio APIs remains the most cost-effective path to production agents.
Source: 36Kr / IT Times — Qwen Office E-Commerce Edition launches (Oct 14, 2026).
On September 28, 2026, Alibaba announced a deep integration between the Qwen App and Quark Cloud Drive. After authorizing, users can type "@Quark Cloud Drive" in a Qwen conversation to search, organize and read documents, photos and videos stored in the drive, then generate study aids, work summaries and photo collections โ turning the drive into a shared "data asset center for users and AI". Files generated by Qwen can also be auto-uploaded to the drive and shared with teams, letting multiple agents collaborate on the same material. Membership benefits are unified too: buying Qwen Elite or above now grants an equal-duration Quark membership, and with the education discount both memberships cost as little as ยฅ25/month. For enterprises, Alibaba is stitching "model + office + cloud storage" into one closed data loop โ when building your own agents, wiring your team knowledge base (cloud drive / OSS) as callable agent context is the pragmatic path, and Model Studio (Bailian) with the Qwen model family remains the most direct entry point.
Source: China.com Tech โ Qwen and Quark cloud drive: files become AI-callable knowledge (Sep 28, 2026).
On September 4, Alibaba disclosed first-month data for Qwen Office, its enterprise general-purpose agent (reported by 36Kr and CNR):
| Metric | Figure | Note |
|---|---|---|
| Registered users | 30M+ in 30 days | One of the fastest-growing AI office products |
| Enterprise share | >50% | The segment with real paying power |
| Release cadence | 120 releases/month | ~4 per day, 1,000+ features |
| Global | International version live | Full global rollout |
| Independent eval | #1 overall | 8 mainstream agent products tested |
Qwen Office consolidated Alibaba's earlier agent products and inherited Alibaba Cloud's enterprise customer base. Early adopters include Changan Automobile, Huifu Payon, Transfar Group, and Home Original Chicken (Laoxiangji), spanning energy, automotive, finance, embodied AI, aerospace, and research.
On September 3, Tmall launched "AI Space Station" โ a token top-up center where users buy subscriptions from leading Chinese model vendors like mobile top-ups: Token Plans, Coding Plans, and pay-as-you-go credits, delivered via redemption codes or direct recharge (reported by DoNews and NBD).
For teams outside China, the takeaway is pricing transparency at market scale. But note: third-party stores are not vendor-direct โ dispute resolution differs. For production workloads, official cloud API channels remain the safer path: free credits, enterprise contracts, invoices, and SLA live on the official side.
On August 31, Qwen's creative platform launched Agent Teams with the Wan 3.0 video model, open to all users (reported by Qianjiang Evening News). From a single idea, specialized agents โ screenwriter, director, art, cinematographer, editor โ collaborate through script, storyboard, character design, video generation, and final cut, covering short dramas, comics, marketing ads, and music videos.
The stack is moving fast underneath: on August 17, Qwen Office open-sourced MyContext, infrastructure that turns scattered work data (messages, docs, approvals, emails, knowledge bases) into context agents can call directly โ it passed 3,000 GitHub stars within weeks. And the August 26 Qwen3.8-Flash model, tuned for agentic tasks, roughly doubled per-task generation speed and cut token consumption by 75% in Qwen Office's real office workloads.
Qwen3.8-Flash โ the same model powering Qwen Office โ is available to any company via Alibaba Cloud Model Studio. Official public pricing:
| Billing item | International (USD / M tokens) | China site (RMB / M tokens) |
|---|---|---|
| Input | $0.15 | ยฅ0.8 |
| Output | $0.47 | ยฅ2.7 |
| Cached input | $0.016 | ยฅ0.1 |
| Rate limits | 5,000 RPM / 5M TPM โ enterprise-grade concurrency | |
| Context | 262K native, 1M via YaRN; text + image multimodal | |
Pricing per Alibaba Cloud Model Studio official pages (effective Aug 27, 2026); new accounts typically receive free credits. Always confirm live pricing on the official campaign page.
Assume each employee chats with an internal support/KB agent 30 rounds/day, averaging 2,000 input + 500 output tokens per round, 22 workdays/month:
| Item | Monthly tokens | Est. cost |
|---|---|---|
| Input | ~13.2M | ~$2.0 |
| Output | ~3.3M | ~$1.6 |
| Total API | โ | ~$3.6/month |
| Server (budget ECS for the business layer) | 1ร 2 vCPU / 2 GB | from ~$14/year |
In other words, for a 10-person team the API bill is negligible โ the real cost is workflow design and integration. That's also why cached input ($0.016/M, ~90% off standard input) matters so much for agents: system prompts and knowledge-base prefixes are reused constantly, and higher cache hits mean dramatically lower bills.
The pattern: nobody left the agent in a chat box. Through IM entry points like DingTalk, agents query CRM and ERP systems for customer, inventory, and order data, then produce business recommendations โ a sales rep sends one message and the agent fetches, organizes, and advises across systems. That's exactly the problem MyContext solves: turning organizational data scattered across systems into context agents understand and can call.
| Dimension | Off-the-shelf agent SaaS | Build on Model Studio API |
|---|---|---|
| Time to value | Instant | 1โ2 weeks of integration |
| Data ownership | Lives with the SaaS vendor | Stays in your own cloud account |
| Customization | Limited by product features | Full control of flows, KB, tool calls |
| Marginal cost | Per-seat subscriptions | Per-token; cached input $0.016/M |
| Compliance | Varies by vendor | Enterprise contracts, invoices, SLA |
| Best for | Individuals, small teams, standard needs | Teams with dev capacity, sensitive data, complex flows |
On September 7, 2026, a four-month-old startup called Mostik (15 people, 12 PhDs, chief scientist Stanislav Smirnov โ 2010 Fields Medalist) demonstrated that a 4B-parameter Qwen-3.5 model running on a phone, paired with a 753B GLM-5.2 model in the cloud, can match GPT-6 Astra on the ARC-AGI 3 benchmark. The bridge transmits only the large model's internal hidden states, never any output tokens, and the small model writes the final answer; both base models are frozen, with the bridge being the only trainable component. Reported gains: the 4B model closes roughly 50% of its performance gap with the 753B model; on the hardest ARC-AGI 3 subsets the 4B's own accuracy improves by up to 2x; and the large model's inference cost falls to ~1/20 of self-generation.
For Qwen Office and any enterprise agent strategy built on Alibaba's model stack, the implication is clear: Qwen already has a credible end-side โ cloud-side collaboration path. Qwen Office can default to a 4B Qwen-3.5 on-device entry point, and dynamically escalate to Qwen 3.8-Max or Qwen3.8-2.4T in the cloud when the task demands deeper reasoning โ cutting inference cost while broadening the range of tasks the agent can take on.
Source: QbitAI โ Fields Medalist joins LLM race: 4B phone-side Qwen + cloud-side GLM dominate ARC-AGI 3 (Sept 7, 2026). The original cost estimate and scenario breakdown above are unchanged; this card only adds the device-cloud collaboration path as a forward-looking reference.
Qwen Office is Alibaba's finished enterprise agent application (SaaS). Model Studio (Bailian) is Alibaba Cloud's model-serving platform providing Qwen-family APIs. You can use Qwen Office directly, or build custom agents on the same underlying models (e.g., Qwen3.8-Flash) via Model Studio APIs.
No. When you call Model Studio APIs, inference runs on Alibaba Cloud's side; your server only handles business logic, the knowledge base, and integrations โ a budget ECS instance is enough. GPU instances (gn6i/gn7i families) matter only if you self-host open-weight models with strict data-residency requirements.
Roughly 750,000 English words of input per dollar of input spend. Feeding a 10,000-word document to your agent costs well under a cent.
Marketplace day-passes are fine for individuals kicking the tires. For production, official cloud channels give free credits, enterprise invoicing, SLA-backed limits (5,000 RPM), and a clean data chain.
Free credits for new users + budget ECS from ~$14/year + Qwen open-weight models available
Claim Free Credits โ View ECS Plans