"Avoid adding a system prompt; all instructions should be contained within the user prompt."
That sentence is on DeepSeek's model card for R1, and as of 26 September 2026 it is still there. It is also the single most-repeated piece of advice about prompting Chinese models, quoted on forums and in Chinese explainers that rank today alongside a companion rule to set the temperature to 0.6. And DeepSeek's own current documentation contradicts it: the worked example on its structured-output page opens with a system message.
Both sentences are real. They are about different models, eighteen months apart, and nobody reconciles them for you. So we went and read what all eight of the major Chinese model vendors actually publish — DeepSeek, Alibaba's Qwen, Zhipu GLM, Moonshot's Kimi, ByteDance's Doubao, Baidu's ERNIE, Tencent's Hunyuan and MiniMax — and filled in a grid of fourteen prompting practices against each one, quoting the vendor's own page for every cell.
The short answer to the question in the title is no. Not one of the eight publishes a rule about how to phrase a prompt in Chinese, or says that a prompt should be written in Chinese at all. What has changed is something else entirely, and it is bigger: seven of the eight now run their flagship model with reasoning switched on by default, and that retires most of the knobs the folklore is about.
A word on two terms before the dossiers. A system prompt is a separate instruction sent alongside your question, conventionally used to set a role or standing rules. Reasoning mode, which these vendors call 深度思考 or thinking, is the model working through a problem privately before answering; when it is on, several of the dials you used to set stop doing anything.
DeepSeek: the rule everybody quotes, and the model it was written for
The "avoid a system prompt" line belongs to DeepSeek-R1, the reasoning model released in January 2025. The same passage tells you to set temperature between 0.5 and 0.7, and to avoid few-shot examples — showing the model a couple of worked cases before your real question — because they degrade its performance. That was a real characteristic of that model.
DeepSeek's current documentation reads differently. Its thinking-mode page states that thinking is enabled by default with the effort set to high, and its structured-output examples pass a system message as a matter of course. The picture is clear enough. The advice was for a specific model; the model moved on, but the advice stayed behind on a page that is still the first result for many searches.
DeepSeek is the only vendor of the eight to publish a warning that few-shot examples hurt a reasoning model. One of eight. If you have been told this is a general property of Chinese models, it is not; it is one company's note about one of its own models.
Baidu: the exception that proves the pattern
Seven of the eight ship reasoning on by default. Baidu is the eighth. On its 千帆 platform you opt in: enable_thinking defaults to false.
Baidu is also alone in publishing a control for making the reasoning shorter — a thinking_strategy with values for a brief think and for chain-of-draft reasoning, alongside a thinking budget that defaults to 16,384 and will not go below 100. No other vendor of the eight offers anything comparable. If your reason for reaching for a Chinese model is that you want the reasoning bounded rather than lavish, this is the one platform whose documentation gives you a lever.
Qwen and GLM: where the switch has stopped existing
Alibaba's documentation splits its own range: the Qwen 3.8 Max series runs hybrid thinking with thinking on by default, while the commercial Qwen 3 series has the same hybrid mode with thinking off. With the same vendor offering opposite defaults on the same documentation page, any general statement like "Qwen does X" falls apart. You have to check the model, not just the brand.
Zhipu goes further. Its thinking parameter defaults to enabled, and on GLM-5.3 and GLM-5.3-Flash it can no longer be switched off at all. Two of the eight platforms now have at least one model where reasoning is not optional. On those, a prompt that says "think step by step" is asking for something already happening.
Which is the one place this guide overlaps something we published before. We tested the prompting folklore across 480 answers and found the tricks did almost nothing. The finding here is not that the tricks fail — it is that on most of these platforms the machine now does the thing the trick was asking for, whether you ask or not.
Kimi, Doubao, Hunyuan and MiniMax: on by default, described four ways
The remaining four all default to reasoning, and the interest is in how differently they say so. Moonshot's Kimi documents thinking as enabled for k2.6 and says k3 "always reasons automatically". ByteDance's Doubao states that models carrying the deep-thinking label have it on when called. Tencent's Hunyuan exposes an EnableThinking flag that is on for its a13b model. MiniMax documents its default as adaptive — the model decides per request.
MiniMax's "adaptive" default is a problem for comparative testing, because you cannot predict from the outside whether a given answer was reasoned or not. If you are comparing two models on cost or latency, that is the variable that will confuse your measurements.
The temperature you set may not be read
Temperature is the dial that governs how much the model varies its wording. Four of the eight state that with thinking on it is ignored, fixed, or explicitly should not be set. Three say nothing either way. Three publish a "change one of temperature or top-p, not both" instruction.
So the honest instruction is not a number. It is: find your vendor's parameter page, check whether temperature does anything on the model you are calling with thinking on, and stop copying 0.6 from a page about a 2025 model. Our guide to the four different length limits one model has makes the same point about a different parameter, and the failure mode is identical — a number that was true of one model on one day, repeated until it looks like a property of the field.
The in-prompt switch is one vendor of eight
The /no_think trick — typing a switch into the prompt to turn reasoning off — is real, is Alibaba's, and is documented for Qwen 3. It is one vendor of the eight. And Alibaba's own model card for the current Qwen 3.6 generation no longer lists it.
If you have been carrying /no_think around as a general technique for Chinese models, it was never general and it is now historical even where it came from.
Language: one vendor says anything at all, and it is not what you would guess
Across all eight vendors' documentation, one says anything about language — and what it says is to name the language you want the answer in, not to write the prompt in Chinese. Two document a language-mixing risk, where a model answers partly in one language and partly in another.
That is the whole of it. No vendor publishes a Chinese-phrasing rule, a Chinese prompt template, or any claim that Chinese input produces better output. If you write in English, or in Bahasa Indonesia or Thai, the thing to do is say which language you want back — because that is the only language instruction any of these companies actually publishes.
Vocabulary is a separate matter and we have already counted it: our guide to the Chinese words these companies use for AI terms measured which word each vendor writes for token, agent, prompt and context length across 730,128 characters of their documentation. Useful for reading their docs; not a rule about how to write to the model.
What to do instead of the folklore
The grid behind this guide has fourteen rows and eight columns, and the cells that matter for an ordinary user come down to four questions you can answer in ten minutes on your vendor's own site.
- Is thinking on for the model I am calling? Assume yes unless you are on Baidu. If it is on, drop "think step by step" — you are asking for something already running.
- Does temperature do anything here? Four of eight say it does not, with thinking on. Check before you tune it.
- Can thinking be switched off? On two platforms, for some models, no.
- What language do I want the answer in? Say so. It is the only language instruction any of the eight publishes.
Our prompt builder lays out the structured fields four of the eight vendors publish frameworks for — task, context, audience, format, constraints. Only task and context are common to all four frameworks, so treat the rest as useful scaffolding rather than a required shape.
Every quotation here was read on the vendor's own documentation on 25 and 26 September 2026. Each of the fourteen practices was classified by hand based on the vendor's own wording, not by pattern matching; cells we could not verify were excluded from our counts. That date matters more than usual. Three of these pages changed while this guide was being written, two documentation dumps we used a fortnight ago now return an error page, and the model names above — Qwen 3.8, GLM-5.3, kimi-k3 — will date this piece faster than anything else in it. The DeepSeek thinking-mode page and the R1 model card are the two documents to read first if you want to see the contradiction for yourself.
View author profile → · Editorial policy
