Where the value range of domestic models (Qwen, GLM, Kimi, MiniMax) actually sits

Establish first where the advantage lies

Domestic models usually cost an order of magnitude less than overseas models in the same class. But that price advantage is not across the board, and its boundary is fairly clear.

Where the advantage is obvious:

• Understanding and generating Chinese content, a corpus advantage that shows up in wording and judgement of register;
• Structured extraction, where the value is high and the quality sufficient;
• High-concurrency, high-volume work, thanks to low price and low latency;
• Direct connection within mainland China, with no extra network setup and stable latency.

Where the advantage is not obvious:

• Complex code refactoring, where the gap in complex engineering contexts is noticeable;
• Scenarios needing very long context, since the ceilings and stability differ across vendors;
• Heavy agent workloads, where tool-calling maturity varies considerably.

So the right approach is to divide work by scenario, not to replace everything.

Selection by scenario

Chinese content work

This is the traditional strength of domestic models, and the advantage is not only price:

• More natural wording: Chinese writing, marketing copy and formal documents read closer to native expression.
• Better contextual understanding: implicit meaning, industry jargon and internet slang are handled better.
• Better format compliance: tables, numbering and hierarchy in Chinese contexts are usually followed more reliably.

Recommendation: for Chinese content generation, test domestic models first. Many teams find the quality no worse than overseas models at a fraction of the cost.

Long text

Domestic models have improved quickly on long context, but note that ceilings differ a lot across vendors and versions, and the versions a relay site carries may lag the official release.

What to do: take the actual list from /v1/models as authoritative, and run a step test for the length you need. See our article on context length and truncation.

Code

This is where domestic models and the leading overseas models differ most clearly. The gaps appear mainly in:

• Large engineering contexts: cross-file, cross-module refactoring and comprehension.
• Edge cases: generated code correct on the happy path and incomplete on error paths.
• Framework conventions: how well specific API details of popular frameworks are known.

Recommendation: everyday code completion and single-function generation are safe to use. For large refactors and critical logic, prefer a stronger model. Or use a two-stage pattern of "domestic model generates, stronger model reviews", which balances cost and quality.

High-concurrency batch work

This is where domestic models offer the best value:

• Low price, so batch costs are predictable.
• Direct connection within mainland China, with stable latency unaffected by cross-border link variation.
• Relatively ample channel supply on relay sites.

Recommendation: make them the default for batch work such as data cleaning, content generation and bulk labelling.

Three things to watch when reaching domestic models through a relay

1. Version lag

This is the most practical problem. Domestic models iterate quickly, but the versions a relay site carries may lag. It shows up as:

• you think you are on the latest version and you are on an older one;
• a name in /v1/models does not match the actual backend.

How to verify: ask it something only the new version can answer correctly, such as a capability or piece of knowledge introduced in that version. That is the most direct criterion.

2. Protocol shape

Domestic models have different native interfaces, but relay sites usually unify them into OpenAI-compatible format, so integration cost is low.

What to watch is regression: once unified into OpenAI format, some native capabilities, such as a particular structured output mode or a particular multimodal shape, may be lost.

If a capability is critical to you, test it once through the native interface and once through the OpenAI-compatible one, and compare.

3. Billing basis

Domestic models are sometimes officially priced per thousand tokens in local currency, while overseas models are priced in US dollars per million tokens. A relay's "multiplier" has a different baseline in each case.

What to do: do not compare multipliers. Compare the amount actually paid per million tokens. That quantity is comparable across the two systems.

An executable combination strategy

The typical effect: most call volume is carried by low-price tiers and only a small number of critical calls go to the expensive tier, bringing total cost down to a fraction of an all-premium setup, with quality loss concentrated in a few stages.

One last reminder: any combination strategy has to be tested against your own real tasks. These four vendors each have different strengths, and a general consensus online cannot substitute for measuring against your own scenario, especially when your tasks have unusual format requirements.