Chinese language models are no longer a single bet. DeepSeek is strong in coding and agents. Qwen has a broad model family. Kimi and MiMo push long-context work. GLM, Doubao, ERNIE, MiniMax and StepFun each cover different production needs.
The hard part is often operational rather than technical. Each provider has its own account, billing flow, model names, rate limits and documentation. A direct account can be the right choice for a large, stable workload. A compatible gateway can be the better choice when a team needs to test several providers without rebuilding its integration.
This guide gives you a practical map.
The provider landscape
| Provider | Useful starting model on AIWave | AIWave input / cache / output per 1M tokens | Context note |
|---|---|---|---|
| DeepSeek | deepseek-v4-pro | $1.914 / $0.064 / $5.742 | Official V4 Pro supports a 1M-token context window. |
| ByteDance Doubao | doubao-seed-2-1-pro-260628 | $1.34 / n/a / $6.70 | Limits depend on the exact Ark model and endpoint. |
| Baidu ERNIE | ernie-5.0 | $2.47 / $0.25 / $9.32 | Check Qianfan's model-specific limit before migration. |
| Z.ai GLM | glm-5.1 | $2.10 / $0.68 / $6.60 | GLM-5.1 documents a 200K context window. |
| Moonshot Kimi | kimi-k3 | $4.50 / $0.90 / $22.50 | Context varies by Kimi model; verify the selected model card. |
| Xiaomi MiMo | xiaomi/mimo-v2.5-pro | $1.56 / n/a / $4.69 | MiMo V2.5 Pro documentation describes up to 1M context. |
| MiniMax | MiniMax-M3 | $0.91 / $0.18 / $3.62 | Check the current MiniMax model page for limits and modalities. |
| Alibaba Qwen | qwen3.8-max | $2.68 / n/a / $8.03 | Model Studio publishes per-model context and regional pricing. |
| StepFun | step-3.7-flash | $0.40 / $0.08 / $2.40 | StepFun maintains a detailed model price table. |
These are AIWave gateway rates dated 2026-08-27, rounded for readability and shown before any account-group multiplier. They are not the providers' direct prices. Rates change; check the dated public pricing endpoint before budgeting.
Direct account or gateway?
Use a direct provider account when one model is already proven, procurement is comfortable with the provider, and volume is high enough to justify separate operations.
Use a gateway when the workload may move between models, the team needs one OpenAI-compatible interface, or separate payment and reconciliation work is slowing delivery. The useful unit is not “number of models.” It is how quickly you can make a tested switch when a model, route or policy changes.
AIWave sits in that second category. It provides one API key, USD billing and a per-request ledger for Chinese-model workloads. Your application still chooses an explicit model ID. The gateway handles the route and records the usage.
A runnable first request
curl https://aiwave.live/v1/chat/completions \
-H "Authorization: Bearer $AIWAVE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"deepseek-v4-pro","messages":[{"role":"user","content":"Return one deployment risk."}]}'If your code already uses the OpenAI SDK, the migration is usually two configuration changes:
client = OpenAI(api_key=os.environ["AIWAVE_API_KEY"], base_url="https://aiwave.live/v1")
response = client.chat.completions.create(model="deepseek-v4-pro", messages=messages)Keep model IDs explicit in configuration. Do not silently swap models in production: output style, tool behavior, tokenization and context handling can differ even behind a compatible API.
How to compare cost honestly
Start with a real request sample, not a headline rate. Export input tokens, cached input, output tokens, retries and failed calls. Then compare the same workload across routes.
For a request with uncached input I, cached input C and output O, the basic estimate is:
cost = I × input_rate + C × cache_rate + O × output_rate
All token counts must use the same per-million unit. Add operational costs separately: payment friction, account administration, engineering time, incident handling and the cost of keeping a fallback route ready.
A gateway may have a higher token rate than a direct account and still cost less to operate. It can also cost more. The answer depends on volume and how much routing complexity your team would otherwise own.
Context windows need engineering discipline
A large context limit does not mean every request should be large. Long prompts increase latency and make failures more expensive. Treat a long-context model as capacity, not a target.
Use chunk-level hashes for stable documents, keep a prompt budget, measure cache hits, and store the provider-reported usage object. When the request approaches a model limit, leave room for output and tool messages. For agent workloads, summarize old turns and preserve only the state that changes the next decision.
A sensible evaluation sequence
- Pick one production-shaped task and define a pass/fail rubric.
- Run it against two or three explicit model IDs.
- Record quality, latency, token usage, errors and retries.
- Price the observed workload using a dated rate card.
- Keep one fallback route, but switch only through a controlled configuration change.
The best provider is the one that survives this test for your workload. If the winner remains stable and volume grows, a direct contract may make sense. If your model choice is still moving, a compatible gateway keeps the experiment readable and reversible.
Primary references
Test the route with your own workload
Use one key, choose an explicit model ID, and check every request against the ledger.
Run a first request