Chinese AI: 7 labs, their models and API prices side by side

In brief

The main Chinese AI labs are DeepSeek, Alibaba's Qwen, Moonshot's Kimi, Z.ai's GLM and MiniMax, plus ByteDance and Tencent. As of November 2026 the paid API prices we could read run from $0.15 per million input tokens for Qwen3.8 Flash to $1.50 for Qwen4 Max, and Z.ai lists GLM-4.7-Flash as free.

What prompt compares Chinese AI models?

Paste this into Whizi, or into any chat that lets you pick a model, once per model you want to test, and fill the brackets.

Seven small glass spheres in different colors on a row of black plinths, with a fine thread linking each to one open notebook, in a charcoal studio under white rim light.

I am testing AI models from [lab, such as DeepSeek, Qwen or Kimi]. Task: [paste one real task]. Answer it in [length]. Then list every fact, number or name you used and mark each as certain, unsure or guessed. End with one question you would ask me before doing this for real.

The last two steps make the model show its gaps, which is what you'll compare. A model that marks nothing as guessed isn't more accurate. It's only quieter. Run the identical text on a model from another lab and compare the marks.

Which Chinese AI models exist?

Seven labs matter for a Western reader as of November 2026: DeepSeek, Alibaba (Qwen), Moonshot (Kimi), Z.ai (GLM), MiniMax, ByteDance (Seed) and Tencent (Hy). A few others, such as StepFun, ship models too. Each one gets its own paragraph here, taken from what the lab itself publishes.

DeepSeek's API docs list two current names, deepseek-flash, served as V4.1 Flash, and deepseek-v4-pro. Both carry a 1M token context and up to 384K tokens of output. Alibaba's Model Studio lists qwen4-max for complex reasoning and coding, qwen3.7-plus for balance and qwen3.8-flash for speed and cost.

Moonshot's Kimi docs name Kimi K3.1 Pro and Kimi K3, and say "All K3 series models support a 1,048,576-token context window." The same page says K2 models are being retired by the end of December 2026. Z.ai's pricing page runs from GLM-5.3 down to GLM-4.7-Flash, which it lists as free. MiniMax's pay as you go page names M3 and M3.1 Flash.

ByteDance's Seed site leads with media models in November 2026, Seedream 5.5 for images, Seedance 2.5 for video and Seed Audio 1.0. It did not mention Doubao by name when we read it. Search coverage describes Doubao as ByteDance's chatbot, but that is a news claim, not the lab's own.

How do Chinese AI models compare on price and context?

They are cheap next to Western flagships, and the range inside the group is about tenfold. The table uses each lab's own price page, read in November 2026. Prices are US dollars per million tokens, sent in and written out. A cell we could not read says not stated.

LabModels on its own docsContextAPI price in / outName in Whizi
DeepSeekDeepSeek V4.1 Flash, V4 Pro1MFlash $0.30 / $1.20 at peak, half off peak. Pro $1.32 / $3.96 at peakDeepSeek V4.1 Flash, V4 Pro
Alibaba Qwenqwen4-max, qwen3.7-plus, qwen3.8-flashUp to 1M per requestMax $1.5 / $4.5. Plus $0.4 / $1.6 to 256K, $1.2 / $4.8 above it (the page also shows a limited time 20% off label). Flash $0.15 / $0.47Qwen4 Max, Qwen3.8 Flash
Moonshot KimiKimi K3.1 Pro, Kimi K31,048,576 (K2 series: 262,144)Not stated (the page tables did not render)Kimi K3.1 Pro, K3
Z.ai GLMGLM-5.3, GLM-5.3-Flash, GLM-4.7-FlashGLM-5.3: 1M, 128K max output. Others not statedGLM-5.3 $1.40 / $4.40. Flash $0.15 / $0.50. GLM-4.7-Flash freeGLM 5.3
MiniMaxMiniMax-M3, M3.1 FlashNot statedM3 $0.30 / $1.20 up to 512K input (a permanent 50% off label; original $0.60 / $2.40). Flash $0.15 / $0.60MiniMax M3, M3.1 Flash
ByteDance SeedSeedream, Seedance, Seed AudioNot statedNot statedSeed 2.0 family
Tencent HyNot readNot statedNot statedHy4, Hy4 flash

Whizi loses on the plain price. Calling a lab's API directly costs those figures, and Whizi sells a plan instead, so a heavy developer who only needs one model won't save money here. Most of these models also sit on Powerhouse, not Pro. DeepSeek V4.1 Flash and Kimi K3 are on Pro.

A worked example shows the spread. Say a job sends 100,000 tokens in and gets 5,000 back. On Qwen3.8 Flash that costs $0.015 plus $0.00235, about $0.017. On Qwen4 Max it costs $0.15 plus $0.0225, about $0.17. That's ten times the bill for the same job, so ask whether the job needs the top model before you pick it. The arithmetic is ours, from Alibaba's list prices.

The price page you read matters. DeepSeek halves its rates off peak, and its peak hours are weekdays 01:00 to 04:00 and 06:00 to 10:00 UTC. MiniMax lists a priority tier at 1.5 times standard. Qwen charges the higher rate for every token in a request once input passes 256K, for qwen3.7-plus.

Three worked examples with Chinese AI models

Each example below is a prompt and the checks a good answer must pass. They are written for this page, not the output of any named model, and we did not run them.

  1. A bug. Prompt: "This Python function returns the wrong total when the list is empty: [paste]. Find the cause, give the fixed function and one test that fails before and passes after." A good answer names the exact line, returns zero for an empty list and includes a test you can run. Try it on DeepSeek and Qwen and compare the tests.
  1. A Chinese document. Prompt: "Summarize this Chinese news article in English in five bullets, and keep every name and number exactly as written: [paste]. Flag any term you translated by guessing." A good answer keeps the figures unchanged, gives names in pinyin with the characters and marks guessed terms. Try Kimi and GLM.
  1. A long report. Prompt: "Here is a 200 page report: [attach]. List every deadline in it with the page number, then quote the sentence that sets each one." A good answer gives real page numbers, quotes sentences that exist in the file and says so if there are no deadlines. Long context is where Kimi's 1,048,576 tokens and DeepSeek's 1M matter.

12 prompts to test any Chinese AI model

Use these in the same words on each model. They are grouped by job.

Reasoning and math.

  1. "A train leaves at 9:40 and travels 212 km at 64 km an hour. When does it arrive? Show each step, then check your answer a second way."
  1. "Here is my working for this problem: [paste]. Find the first wrong step and stop there."

Code.

  1. "Write a [language] function that [does X]. List the edge cases, then write a test for each."
  1. "Review this code for bugs and security problems: [paste]. Rank them from worst to least bad and give a fix for the worst one."
  1. "Explain what this error means and the two most likely causes: [paste error and the code around it]."

Writing and translation.

  1. "Translate this into [language] for a [formal or casual] audience, and give two alternatives for any phrase that has more than one natural reading: [paste]."
  1. "Rewrite this email in under 90 words and keep the tone [tone]: [paste]."

Research and documents.

  1. "Using only the text below, answer [question]. Quote the sentence you used. If the text does not say, reply not stated: [paste]."
  1. "Search the web and give me the current price of [product] from the vendor's own page, with the link and the date you read it."
  1. "Make a table comparing [A] and [B] on [three criteria], using only each one's own page, and mark any gap as not stated."

Judgment calls.

  1. "Give the strongest case for and against [claim], three sentences each, with a source I can open for each side."
  1. "Answer this twice, once as briefly as you can and once in full, then tell me what the short version left out: [question]."

How do you get a good result from a Chinese AI model?

Test on your own task and judge before you read. The five steps above do that. Four mistakes spoil most comparisons, and each has a fix.

  1. Changing the prompt between models. Fix: paste the identical text each time.
  2. Trusting a benchmark over your task. Fix: use public scores to shortlist, then decide on a job you will really do. Qwen vs DeepSeek shows what the Artificial Analysis scores look like next to cost per task.
  3. Pasting private material into an app without reading its data policy. Fix: read the lab's own policy first. Is DeepSeek safe explains the difference between the app and the model.
  4. Reading the cheapest sticker price as the cost. Fix: add peak hours, tiers and long input surcharges, which the table above lists.

Is Chinese AI safe to use?

It depends on the app, not the nationality of the model. A lab's own app sends your chats to that lab, under that lab's privacy policy and legal system. Read the policy for any app before you paste anything private. Our DeepSeek safety page and Qwen free page show where each of those two apps says it stores data.

A model reached through another provider is governed by that provider's policy instead. That is how the models in Whizi work: you send the message to Whizi, and Whizi routes it. We did not test how any of these models answer politically sensitive questions, so we make no claim about it. If that matters to your use, ask the same question of a Chinese model and a Western one and compare.

Where does Whizi fit for Chinese AI?

Whizi is one app instead of a pile of single purpose tools. Its picker lists models from DeepSeek, Qwen, Kimi, GLM, MiniMax, Tencent, StepFun and ByteDance next to Claude, GPT and Gemini. One prompt can go to two of them side by side. We confirmed each name above in the live model list in November 2026. Side by side Compare is on Pro and Powerhouse.

The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card, and it has no Chinese model. DeepSeek V4.1 Flash and Kimi K3 start on Pro, whose first week is $1.99, and most of the others sit on Powerhouse. Web search is on by default in a new chat. You can upload a PDF, Word or Excel file, and a scanned PDF gets a vision transcription pass.

Whizi loses if you only want one model through its own free app, or the lowest direct API price. The lab's own site is cheaper then. It also lacks anything a lab ships only in its own app.

What can these models not do for you?

They can invent facts, like every language model, and a cheap price does not change that. Why AI gets things wrong covers the mechanism, and the answer is the same here: open the source behind any number you plan to use.

Long context doesn't mean good recall. A 1M token window lets you paste a book, but it says nothing about whether the model finds the one clause you need. Test it: hide a known fact in the middle of a long file, ask for it with a quote, and see whether the model finds it or invents one. That check takes five minutes and beats any spec sheet.

Model names and prices on these pages change fast. Kimi's docs already say K2 is being retired by the end of December 2026, and DeepSeek renames models as it ships. Check the lab's page on the day you build something. If you want Chinese for learning, not for model choice, read best AI for Chinese.

Workflow checklist
  • Write three pass or fail checks before you read any answer
  • Send the identical prompt to every model you test
  • Read the lab's own price page on the day you decide
  • Read the privacy policy before pasting private material
  • Open the source behind any number you will reuse
Common questions

Frequently asked questions

What is the best Chinese AI?

No single one wins every job, and we ran no test to rank them. On the labs' own price lists, Qwen4 Max is the dearest flagship at $1.50 per million input tokens and Qwen3.8 Flash the cheapest at $0.15, as of November 2026. Test two on your own task.

Is DeepSeek the only Chinese AI?

No. Alibaba's Qwen, Moonshot's Kimi, Z.ai's GLM, MiniMax, ByteDance and Tencent all ship models, and each one is named in the table above with the lab's own price page as the source.

Is there a free Chinese AI?

Z.ai's price page lists GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash as free for API use, as of November 2026. Other labs run free apps whose limits we did not read, so check their own pages.

Can I use Chinese AI models outside China?

Yes. DeepSeek, Alibaba Cloud, Moonshot, Z.ai and MiniMax each publish an international API price list in US dollars, and models from these labs are in the Whizi picker. Region rules for each app are on its own site.