Alibaba Users Manual

The current Qwen lineup

As of 2026-05-05, Qwen 3.6 Max is Alibaba's flagship for complex agentic workflows. The lineup splits into proprietary (Max, Plus) and open-weights (3.6 / 3.5 / Omni) tiers, plus dedicated multimodal models (Wan, HappyHorse, Image).

Qwen text/reasoning lineup

Qwen 3.6 Max — deep dive

This is the model the rest of the manual is centered on. Qwen 3.6 Max is Alibaba's flagship as of April/May 2026, designed end-to-end for autonomous agent workflows.

Qwen 3.6 Max vs the competition (positioning)

Release timeline (chronological)

How to pick a Qwen model

Pricing

Alibaba Cloud Model Studio is pay-as-you-go in USD or CNY depending on the region (China, Singapore, international). Headline characteristics of the Qwen pricing structure:

• Qwen 3.5 Series shipped with a 60% price cut over the prior generation, and Qwen 3.6 continues that trajectory.
• Qwen 3.6 Plus is positioned as the cost-efficient tier — the explicit "match larger predecessors at smaller cost" model.
• Open-weights variants (Qwen 3.6 / 3.5) have zero per-token cost when self-hosted; pay only for compute.

Open-weights variants

Alibaba is one of the most active open-weights labs in the world. The current open Qwen family at huggingface.co/Qwen includes:

All-modalities overview

Wan 2.7 (video generation)

Wan is Alibaba's video-generation line. Wan 2.7 is the latest entry in Model Studio (April 2026).

• Modalities: Text-to-video.
• Surface: Available via Alibaba Cloud Model Studio.
• Use cases: Marketing creative, product demos, education, social-media content.
• Compared to: Veo 3 (Google) is more polished cinematically; Grok Imagine Video is cheaper. Wan 2.7 is the China-native option with first-party Alibaba Cloud distribution.

HappyHorse 1.0 (image-to-video)

Released April 2026. Distinct from Wan: HappyHorse takes an existing image as input and animates it.

• Modality: Image-to-video — start from a still, get a video out.
• Strength: High-fidelity, realistic dynamic rendering. Top-ranked among image-to-video models per Alibaba's release notes.
• Use cases: Animating product photography, bringing static art to life, character animation.
• Alternative use: Pair with Qwen Image (or any image gen) to go text → image → video in two steps.

Qwen Image (text-to-image)

Open-weights image-generation family on Hugging Face. The current generation is Qwen Image 2512.

• Modality: Text-to-image.
• Distribution: Open weights on Hugging Face; available via Model Studio API.
• Strengths: Strong on Chinese-language prompts; competent on photorealism and stylized output.
• Pair with: HappyHorse 1.0 to animate the output, or feed into Wan 2.7 for video continuation.

Qwen 3.5 Omni (audio + multimodal)

The "everything in one model" tier. Qwen 3.5 Omni is the only Qwen variant that natively handles audio input/output alongside the standard text + image + video.

• Modalities: Text, audio, image, video — all in one open-weights model.
• Distribution: Open weights on Hugging Face; demos available online and offline.
• Use cases: Voice agents that can also see, multimodal accessibility, cross-modal search ("find the moment in this video where someone says X").
• Tradeoff: Generalist — for hardest-tier text reasoning, Qwen 3.6 Max wins; Omni's edge is breadth across modalities.

Practical workflows

Three end-to-end pipelines you can copy. Each names which Qwen model handles which step.

Workflow A — Marketing video (15s) from a product brief

1. Qwen 3.6 Max — plan

   Send the product brief and target audience. Ask for: 3 still-frame prompts, motion prompts for each, voiceover script with timestamps. (Use the "Marketing video pipeline" template under Use-case library.)

2. Qwen Image — generate stills

   Run each still-frame prompt to produce 3 hero frames at high resolution.

3. HappyHorse 1.0 — animate each still

   Send each still + its motion prompt. Get 3 short animated clips.

4. Wan 2.7 — optional bridge shots

   If you need transitions or B-roll between hero shots, generate them text-to-video.

5. Off-Alibaba — assemble

   Stitch in any video editor; layer the voiceover. Total runtime 15s; total Qwen calls ~7 if you don't iterate.

Workflow B — Animate an existing product photo

1. Qwen 3.6 Max — direction

   Upload the product photo. Use the "Image-to-video direction" template to get a tuned motion prompt for HappyHorse.

2. HappyHorse 1.0 — render

   Send the photo + motion prompt; get the looped clip back.

3. Qwen 3.6 Max — review

   Optionally send screenshots from the rendered clip back to Max and ask for QC notes — does the motion match brand guidelines? Anything to revise?

Workflow C — Multilingual product launch creative

1. Qwen 3.6 Max — base creative

   Generate the headline, body copy, and CTA in your source language.

2. Qwen 3.6 Max — localize

   Use the "Localize marketing copy" template (under Use-case library) for each target market — it goes beyond literal translation.

3. Qwen Image — localized stills

   For markets where image text matters, regenerate stills with localized typography. (Note: text-in-image rendering is improving but verify each output.)

4. HappyHorse / Wan — localized motion

   Animate the localized stills. The motion prompt itself usually doesn't need localization.

chat.qwen.ai — setup

1. Visit chat.qwen.ai and sign in (Qwen / Alibaba account).
2. Default model is the latest Qwen flagship. The chat surface routes to the current best Qwen automatically; specific model selection (Max vs Plus vs others) may be available via a model picker depending on region.
3. Toggles to know:
   • Thinking — turn on chain-of-thought reasoning for hard tasks.
   • Web search — ground responses in current web results.
   • File upload — drop in PDFs, code, docs, images. With long context you can upload sizable files.
   • Image / Video tools — generate images via Qwen Image, video via Wan or HappyHorse, depending on the surface area of the chat experience in your region.
4. Rate limits apply. For guaranteed throughput or production access, use Model Studio API.

Modes & tools — when to use each

Optimal prompts for chat.qwen.ai

Qwen models reward structure. Below are tested templates for the chat surface.

Account & keys

1. Visit Alibaba Cloud Model Studio and sign in with your Alibaba Cloud account. Choose the region that matches your latency / data-residency needs (China, Singapore, international).
2. Add a payment method. Pay-as-you-go in CNY or USD depending on region.
3. Generate an API key from the dashboard. Treat as a password — store in env vars, not in code.
4. Verify access: hit the chat completions endpoint with a small test request to confirm the key and region.

First API call

The Model Studio chat completions endpoint is OpenAI-compatible. If you've used the OpenAI SDK, the only changes are base_url and model.

OpenAI-compatible vs DashScope native

Function calls & JSON

Qwen 3.6 Max supports OpenAI-style tool calls and structured output. This is the foundation for agentic workflows.

Agentic tasks (Qwen 3.6 Max's flagship use case)

Qwen 3.6 Max is explicitly tuned for autonomous agent workflows. The minimum viable loop:

Self-host (open weights)

Qwen 3.6 (open variants) and Qwen 3.5 Omni / 397B are downloadable from Hugging Face. Common deployment paths:

• Ollama for local laptop dev.
• vLLM for production GPU serving.
• Text Generation Inference (TGI) for Hugging Face-native serving.
• llama.cpp for quantized inference on commodity hardware.
• Alibaba Cloud PAI / EAS for first-party self-managed serving on Alibaba infra.

Streaming responses

For chat UIs and long generations, stream responses so users see progress instead of waiting for the full payload. The OpenAI-compatible endpoint supports stream=true exactly like the OpenAI SDK.

Rate limits, retries & errors

Treat the API as eventually reliable, not perfectly reliable. The minimum viable error-handling shape:

• 429 (rate limit): back off with jitter. Start at 1s, double up to ~30s. Don't retry forever — give up after 5 attempts and surface to the caller.
• 5xx (transient): retry up to 3 times with backoff; then fail loud.
• 4xx other than 429: don't retry — these are bugs in your request shape (bad model name, malformed JSON, missing field). Log and fix.
• Timeouts: set a sensible client timeout (e.g. 60s for Max non-streaming, longer for streaming). Don't inherit the default — it's usually too long.

Cost-control patterns

• Two-tier routing. Send simple work (classification, extraction, formatting) to qwen3.6-plus; reserve qwen3.6-max for actual reasoning / agentic / visual tasks. Often cuts cost 60-80% with no quality drop.
• Cap output tokens. Set max_tokens deliberately. With long-context inputs it's tempting to let output run unbounded — costs add up fast.
• Short system prompts. A 5-line system prompt that defines role and constraints crisply outperforms a 30-line one in most evals — and costs less per call.
• Stop iterating in chat for large workflows. Convert "tell me more about X" loops into one well-scoped prompt — each iteration replays the entire context.
• Choose the right region. Latency and pricing differ across China / Singapore / international Model Studio regions. Match to where your users (and data) live.

Multi-region considerations

Alibaba Cloud Model Studio runs in multiple regions. The same model ID (e.g. qwen3.6-max) can have different latency, availability, and per-1M-token pricing depending on the region your endpoint lives in. Practical rules:

• Pin a region per environment (dev / staging / prod). Don't let it drift via DNS.
• Test at least one alternate region so you have a documented failover path.
• Data-residency-sensitive workloads: the Singapore-international region is typically the right choice for non-PRC data flows; verify the current terms before committing.
• Don't assume PII-handling parity across regions. The contracts and certifications differ.

Use-case library

Tested templates organized by what you're trying to accomplish. Each one assumes Qwen 3.6 Max unless noted; drop down to qwen3.6-plus when the task is simpler and cost matters.

Agentic app development (Qwen 3.6 Max's headline)

Visual browsing & UI reasoning (the other Max headline)

Long-context analysis (1M+ tokens)

Coding & engineering

Multimodal pipelines (Qwen Image / Wan / HappyHorse)

Data extraction & structured output

Translation & multilingual (Qwen's strength)

Research & decision-making

Writing & editing

Prompt builder

Patterns library

Reusable shapes that improve quality independent of the specific task. Most apply to any model; a few are tuned to Qwen 3.6 Max specifically.

The "plan-execute-verify" pattern (Qwen 3.6 Max)

For any non-trivial agentic task, ask Max to output a plan first, wait for approval, then execute. Catches misunderstandings early and uses Max's tuning correctly.

The "two-tier" pattern (Plus → Max)

For multi-step pipelines, route simple steps to Qwen 3.6 Plus (cheap, fast) and escalate only the hardest to Max. Often cuts cost 60-80% with no quality loss. Good rule of thumb: triage / classify / route / format → Plus. Reason / plan / write the substantive answer → Max.

The "modality handoff" pattern

Qwen 3.6 Max plans → Qwen Image generates stills → HappyHorse / Wan animates. Each model handles its strongest modality. Single-vendor pipeline keeps integration simple. Bonus: Qwen 3.5 Omni in the loop if audio is involved.

The "schema first" pattern

When you need structured output, define the schema in the prompt before showing the input. Qwen follows declared schemas reliably; ad-hoc "give me JSON" prompts produce more drift.

The "self-critique" pattern

After a substantive answer, ask the model to critique its own output before you accept it. Often catches subtle errors a second pass would have caught.

The "show your work first" pattern

For reasoning-heavy tasks, ask for the work before the answer. Reduces sycophantic / shortcut answers and makes errors easier to spot.

The "bounded context" pattern

1M tokens is a tool, not an obligation. If you're using only a fraction of the context, say so explicitly — it lowers the chance of the model anchoring on irrelevant earlier turns.

The "force the disagreement" pattern

Qwen models can be agreeable. To stress-test a plan, ask it to argue the opposite case before recommending.

Anti-patterns

Things that look helpful but consistently make outputs worse.

Universal rescue prompts

When an answer is off, these one-liners reliably get a better one without needing to rewrite the original prompt.