From clicking a link in a browser to running a trillion-parameter model on your own hardware. Pick the row that matches how much control you need.
The agentic consumer assistant, wired into Taobao, Alipay, Fliggy and Amap. Currently China-focused.
iOS & Android
Alibaba Cloud Model Studio / Qwen Cloud, with OpenAI- and Anthropic-compatible endpoints and 1M free trial tokens.
Download the weights from Hugging Face or ModelScope and serve them with vLLM, SGLang, llama.cpp or Ollama.
Per million tokens, Alibaba Cloud Model Studio international deployment, as of mid-August 2026. Regional pricing and promotional tiers differ — confirm before you budget.
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
| Qwen3.8-Max | $2.00 | $6.00 | 1M | Long-horizon agents, autonomous coding, research work |
| Qwen3.7-Max | $2.50 | $7.50 | 1M | Previous flagship; office automation and tool use |
| Qwen3.5-397B | $0.60 | $3.60 | 128K | Strong general reasoning at mid cost |
| Qwen3.5-Plus | $0.40 | $2.40 | 1M | Balanced default; rises to $0.50 / $3.00 above 256K input |
| Qwen3.5-Flash | $0.10 | $0.40 | 1M | High volume with a real context budget |
| Qwen3.7-Flash | $0.03 | $0.13 | 1M | Cheapest tier: classification, extraction, routing |
Qwen Cloud exposes OpenAI-compatible and Anthropic-compatible interfaces, so most existing client code works after changing the base URL and model name.
from openai import OpenAI
client = OpenAI(
api_key=os.environ["DASHSCOPE_API_KEY"],
base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)
resp = client.chat.completions.create(
model="qwen3.8-max",
messages=[
{"role": "system", "content": "You are a precise research assistant."},
{"role": "user", "content": "Summarise this quarter's filing in five bullets."},
],
# low | medium | xhigh - trades latency and cost against depth
extra_body={"reasoning_effort": "medium"},
)
print(resp.choices[0].message.content)curl https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions \
-H "Authorization: Bearer $DASHSCOPE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.5-flash",
"messages": [{"role": "user", "content": "Explain MoE routing in two sentences."}],
"stream": true
}'Alibaba Cloud advertises 1 million free tokens for new Model Studio users, which is enough to evaluate several models against a real workload before committing.
Qwen models are resold through OpenRouter and most aggregator platforms, and deployable with a few clicks in Alibaba's PAI-EAS if you want a managed endpoint with fine-tuning attached.
Every open Qwen lands on Hugging Face and ModelScope on release day, and the major inference stacks now ship day-zero support.
# A single 27B dense model - the practical default for one GPU box
vllm serve Qwen/Qwen3.8-27B \
--max-model-len 262144 \
--served-model-name qwen3.8-27bfrom transformers import AutoModelForCausalLM, AutoTokenizer
name = "Qwen/Qwen3.8-27B"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name, dtype="auto", device_map="auto")
msgs = [{"role": "user", "content": "Write a regex for ISO-8601 durations."}]
text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
ids = tok([text], return_tensors="pt").to(model.device)
out = model.generate(**ids, max_new_tokens=2048, temperature=1.0,
top_p=0.95, top_k=20)
print(tok.decode(out[0][len(ids.input_ids[0]):], skip_special_tokens=True))Qwen publishes these defaults for the 3.8 line, and drifting from them is a common cause of disappointing local results:
A one-million-token window is a capability, not an instruction. Retrieval into a 32K prompt is usually faster, cheaper and more accurate than dumping a corpus into a 1M window.
Model Studio has separate international and mainland-China deployments with different endpoints, model availability and terms. Pick deliberately if you have data-residency obligations.
Apache 2.0 is common but not universal. The Max-class open release carries its own bespoke licence, and some checkpoints are research-only.
With four generations shipped in six months, pin the exact model ID in production and re-evaluate deliberately rather than tracking a moving alias.
Reasoning output counts as output tokens. On
models with a reasoning_effort dial, tuning it down is often the cheapest real saving.
The gaps between frontier models on public benchmarks are now smaller than the gap between good and bad prompting on your specific task.