Chinese large language models have closed much of the gap over the past two years. DeepSeek is known for price-performance, Alibaba's Qwen for Chinese-language understanding, and Zhipu's GLM for coding and agent work. Using all three directly means three sign-ups, three balances, three SDK integrations and three bills. This guide shows how to call DeepSeek, Qwen and GLM through a single Zivv key, how to choose between them, and how to plug them into existing code and into Claude Code.
Which Models Are Available
These are the main Chinese text models in the Model Hub today (always confirm ids against the live list):
| Vendor | Recommended model | Positioning |
|---|---|---|
| DeepSeek | deepseek-v4-1-flash-261017 | V4.1 Flash: hybrid reasoning, stronger agentic coding and long-horizon execution, low cost |
| DeepSeek | deepseek-v4-pro-ga-260813 | V4 Pro: complex reasoning and coding |
| Alibaba Qwen | qwen3.7-max | Flagship for complex tasks and Chinese understanding |
| Alibaba Qwen | qwen3.7-plus | Balanced capability and cost |
| Zhipu | glm-5.2 | Strong at coding and agent tasks |
All of them have a 1M-token context and support reasoning, tool use and prompt caching. DeepSeek and Qwen also support structured output, and DeepSeek V4.1 Flash adds web search. Each model page lists its exact capabilities.
Choosing by Cost and Task
Start with price. Using Claude Sonnet 5.5 as a reference point, at Model Hub prices when this was written:
- DeepSeek V4.1 Flash: input and output both cost roughly a third of Sonnet 5.5 or less — the default for batch jobs and cost-sensitive work
- Qwen3.7-Plus: cheaper than Sonnet 5.5, a good fit for Chinese-heavy everyday workloads
- DeepSeek V4 Pro and GLM-5.2: input slightly above Sonnet 5.5, output about the same; you pick them for their strengths, not to save money
- Qwen3.7-Max: flagship pricing, more expensive than Sonnet 5.5; choose it for Chinese understanding and hard tasks
Prices follow the upstream providers, so check the Model Hub for current rates. Then match the task:
| Workload | Recommendation |
|---|---|
| High-volume summarization, classification, extraction, translation | DeepSeek V4.1 Flash |
| Chinese customer support, copywriting, formal documents | Qwen3.7-Plus, or Qwen3.7-Max when quality matters most |
| Code generation and agent tasks | GLM-5.2 or DeepSeek V4 Pro |
| Production jobs that must stay reproducible | A dated snapshot version |
This is rarely an either-or decision. A common setup sends hard coding work to Claude and high-volume Chinese text processing to a Chinese model, switching per task inside the same codebase. The multi-model routing guide covers the pattern in depth.
Groups: One Model, Several Upstreams
Every Zivv key belongs to a group. Chinese models are grouped by upstream provider: DeepSeek, for example, is served through both Volcengine and Alibaba Cloud Bailian, each as its own group with its own price multiplier, and GLM also has two. The group prices section on each model page shows which groups carry the model and at what price.
You can use them in two ways:
- Single-group key: one group, so the price and upstream are fixed — useful when billing needs to be unambiguous
- Smart routing key: set the routing mode to smart routing and select several groups; Zivv picks channels from the lowest price multiplier upward, so one key covers DeepSeek, Qwen and GLM
Note that the Claude MAX group only contains Claude models. Calling DeepSeek with a key in that group fails; create a separate key or use smart routing. The 403 troubleshooting guide explains permission errors in detail.
Calling Them with the OpenAI SDK
All of these models speak the OpenAI Chat Completions protocol, so existing OpenAI code only needs a new base_url, api_key and model:
from openai import OpenAI
client = OpenAI(base_url="https://zivv.pro/v1", api_key="sk-your-key-here")
for model in ["deepseek-v4-1-flash-261017", "qwen3.7-plus", "glm-5.2"]:
resp = client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": "Introduce yourself in one sentence."}],
max_tokens=200,
)
print(model, "=>", resp.choices[0].message.content)Streaming and tool calls work exactly as they do with GPT models; see the Python quickstart and the Node.js quickstart. If you are moving an existing project off an official endpoint, the OpenAI-compatible migration checklist lists what to verify.
Calling Them with the Anthropic SDK
The same models also support the Anthropic Messages protocol, so projects built on the Anthropic SDK only change the model:
import anthropic
client = anthropic.Anthropic(base_url="https://zivv.pro", api_key="sk-your-key-here")
msg = client.messages.create(
model="deepseek-v4-pro-ga-260813",
max_tokens=1024,
messages=[{"role": "user", "content": "Explain what an idempotent API is."}],
)
print(msg.content[0].text)The OpenAI Responses API is supported by some versions only: Qwen3.7 and GLM-5.2 support it, while the DeepSeek versions served through Volcengine do not. Before pointing a Responses-based tool such as Codex at one of these models, check "supported protocols" on its model page.
Running Claude Code on a Chinese Model
Because they support the Anthropic protocol, DeepSeek, GLM and Qwen can drive Claude Code directly — handy when you make lots of routine changes and want to keep costs down:
export ANTHROPIC_BASE_URL=https://zivv.pro
export ANTHROPIC_AUTH_TOKEN=sk-your-key-here
export ANTHROPIC_MODEL=deepseek-v4-pro-ga-260813
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-1-flash-261017
claudeANTHROPIC_DEFAULT_HAIKU_MODEL controls the model Claude Code uses for lightweight background work; setting it as well keeps those requests from asking for a Claude model. The key must have access to the group that serves DeepSeek.
One caveat: Claude Code's prompts and tools are tuned for Claude. Chinese models handle routine edits, test writing and code reading well, but for complex cross-file refactors and long agent runs Claude is usually more reliable. A practical setup is to default to a Chinese model and use /model to switch to claude-sonnet-5-5 or claude-opus-5-5 when a task gets hard. The Claude model comparison explains which Claude to pick.
FAQ
Q: Is there any difference compared with using each vendor's own platform? A: Zivv forwards to the vendor or cloud upstream, so the model is the same. The difference is operational: one key, one balance and one usage log for every model, with no separate sign-ups or top-ups.
Q: Should I use the ids with a date suffix, like deepseek-v4-1-flash-261017? A: Dated ids are fixed versions whose behavior does not change with upstream updates, which is what you want in production. Use the undated id if you always want the latest version.
Q: I get "model not found". What should I check? A: Confirm the id matches the Model Hub exactly, then confirm your key's group includes the model. The 404 model-not-found guide has the full checklist.
Q: Can these models read images? A: The recommended versions listed here are text-only. For image understanding, use the Claude, GPT or Gemini families; the Model Hub can filter by input modality.
Wrapping Up
Chinese models shine in high-volume, Chinese-heavy, cost-sensitive work, and combining them with Claude and GPT can cut overall spend substantially. Sign up for Zivv, create a smart routing key, and call DeepSeek, Qwen, GLM and international models from the same code, picking the best model for each task.