Choose a Chinese LLM: Current ChinaAPI Routing Guide

The wrong default is a flagship model for every prompt. Most production failures happen earlier: a text-only model receives an image, a 256K model gets an unbounded repository, a cheap first pass silently becomes five retries, or a tool loop has no budget and approval limit.

The 2026-07-20 gateway-reconciled snapshot contains 25 current per-token models tagged Reasoning. That is enough choice to make a generic leaderboard less useful than a routing rule. This guide narrows the first decision to the constraint that can make a task fail: modality, context window, tool use, latency, or the cost of an accepted answer.

This is a current-catalog selection guide, not a public quality benchmark. The model IDs below are live in the ChinaAPI catalog today; account group, balance, rate limits, upstream health, and the task itself still affect whether a request succeeds.

The short answer

• Start text-first RAG, extraction, classification, and inexpensive code candidates with deepseek-v4-flash. It has a 1M context window and tool support in the current catalog at $0.1389 input / $0.2778 output per million tokens. Do not send image or video inputs to a text-only route.
• Use qwen3.7-plus as the practical general-purpose agent default when the job needs files or images as well as a 1M window. It is the right first test when tool use and visual context matter more than minimizing every token.
• Choose the operation before the brand for long work. LongCat-2.0 is a 1M text-only tool route; kimi-k2.7-code is the code-specialist candidate at 256K; MiniMax-M3 is the 1M route in this snapshot with files and vision. They solve different constraints.
• Reserve qwen3.7-max, glm-5.2, or kimi-k3 for selected high-value passes and independent review. A premium pass is most useful after a candidate exists, not as the first call in a bulk workflow.
• Use glm-5v-turbo when the input includes video. In the current catalog it is the reasoning route explicitly tagged for image, video, files, and tools; a text-only or image-only route cannot recover information it never received.

The position is deliberate: do not route by a vendor's flagship label. Route by the failure mode you cannot accept, then earn a premium escalation with accepted-output data.

What the current catalog makes possible

The table is intentionally a routing map rather than a rank order. deepseek-v4-flash can be a better first call than a flagship when the input is text and the goal is to cheaply produce candidates. It is a bad first call when the job depends on an image, a video, or a file the model cannot inspect.

Price is an input to routing, not the decision

The current display prices show why token direction matters. deepseek-v4-flash is $0.1389 input / $0.2778 output per million tokens. qwen3.7-plus is $0.2778 / $1.1111; LongCat-2.0 is $0.2778 / $1.1111; and qwen3.7-max is $1.6667 / $5.

Those numbers are current gateway-reconciled display rates, not a prediction of total project cost or a quality score. Thinking tokens, cache misses, tool calls, retries, human repair, and the probability of accepting the answer all change the outcome. The useful unit is:

usable-answer cost =
  (input tokens + output/thinking tokens + tool costs + retries + review time)
  / accepted answers

A low input price loses if a model cannot process the source modality or if its first result repeatedly fails a schema, citation, tool, or code-review check. Conversely, a high-output model may be appropriate for one selected final review and wasteful for every candidate in a batch. Check live pricing before putting a route into production.

A repeatable two-stage routing method

Make the task card before choosing the model. The point is to prevent a prompt from accidentally becoming an expensive, unbounded agent loop.

{
  "task_id": "contract-qa-01",
  "task_type": "document_question_answering",
  "source_modalities": ["pdf", "image"],
  "context_band": "1m",
  "needs_tools": true,
  "needs_video_input": false,
  "max_tool_rounds": 6,
  "max_attempts": 2,
  "acceptance_checks": ["answer cites supplied pages", "JSON validates", "no unsupported claim"],
  "candidate_model": "qwen3.7-plus",
  "escalation_model": "glm-5.2",
  "human_approval_required": true
}

Then use two stages:

1. Select the lowest-cost route that can actually read the inputs and meet the context/tool constraints.
2. Run a limited candidate pass with explicit acceptance checks, cost and retry caps.
3. Escalate only the failed or high-value cases to a premium or specialist route.
4. For consequential output, perform an independent review pass and require human approval before external actions, payments, deployments, or data deletion.
5. Store the accepted/rejected result, billed tokens, tool rounds, retries, and repair minutes. Promote a route to default only after it wins on accepted-output data.

The internal routing brief defines a 60-task candidate set spanning text RAG, structured extraction, coding, tool use, long documents, image or video understanding, and final review. That is a test plan, not a public benchmark. We have not yet published repeated ChinaAPI completion rates, latency distributions, tool-success rates, or repository-level coding results across all 25 models.

Common routing mistakes

What this guide does not claim

This article does not publish a universal quality ranking, latency league table, vendor availability guarantee, or safety-policy outcome. Model configurations and gateway prices can change; the live pricing page is authoritative. We also do not yet have a public benchmark that proves one model is better at code, reasoning, tool use, or multimodal understanding for every task.

Use the current catalog facts to choose a defensible first route. Then measure your own accepted outputs, failure types, retries, and total cost. The next useful update is a benchmark report with a fixed task set, repeated runs, raw methodology, and clearly scoped conclusions — not a louder leaderboard.