Check this out: i Tested Four Chinese AI Models and Here's What I Found
When I graduated from bootcamp last year, I figured I'd spend most of my career wiring up OpenAI or Anthropic APIs and calling it a day. Then somebody in a Discord server casually dropped the phrase "DeepSeek" and I realized I had absolutely no idea how much was happening outside the Western AI bubble.
I was shocked. Honestly, I felt behind. So I did what any curious new dev would do — I started digging, running prompts, comparing outputs, and watching my bill. After weeks of testing, I want to share what I learned, in plain English, because the pricing alone made my jaw drop.
This whole piece is my honest take on four model families out of China: DeepSeek, Qwen, Kimi, and GLM. I'll walk you through the ones I tried, what each one is good at, where they fall short, and how you can actually use them through one simple endpoint.
Let's go.
So Who Are These Four, Anyway?
Before I get into my notes, here's the quick lay of the land. These are all Chinese-built large language models, and each comes from a different company:
- DeepSeek comes from a quant trading firm called 高幻方 (literally "High Magic Square"). Weird origin story, amazing models.
- Qwen is built by Alibaba. Yes, that Alibaba. The e-commerce giant. I had no idea they were this deep into AI.
- Kimi is the project from Moonshot AI, which translates to "Dark Side of the Moon." Cool name, scary-smart outputs.
- GLM is from Zhipu AI, one of the earliest Chinese AI labs, founded by folks from Tsinghua University.
When I realized four serious players were all building competitive models, it kind of blew my mind. We're spoiled for choice now.
Quick Cheat Sheet I Wish I'd Had on Day One
Here's the table I keep open in a tab whenever I start a new project. Everything here comes from actual pricing pages and my own testing through Global API's unified endpoint.
| Feature | DeepSeek | Qwen | Kimi | GLM |
|---|---|---|---|---|
| Developer | DeepSeek (幻方) | Alibaba (阿里) | Moonshot AI (月之暗面) | Zhipu AI (智谱) |
| Price Range | $0.25-$2.50/M | $0.01-$3.20/M | $3.00-$3.50/M | $0.01-$1.92/M |
| Best Budget Model | V4 Flash @ $0.25/M | Qwen3-8B @ $0.01/M | N/A (all premium) | GLM-4-9B @ $0.01/M |
| Best Overall | V4 Flash @ $0.25/M | Qwen3-32B @ $0.28/M | K2.5 @ $3.00/M | GLM-5 @ $1.92/M |
| Code Generation | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ |
| Chinese Language | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ |
| English Language | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ |
| Reasoning | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ |
| Speed | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐⭐ |
| Vision/Multimodal | Limited | ✅ (VL, Omni) | ❌ | ✅ (GLM-4.6V) |
| Context Window | Up to 128K | Up to 128K | Up to 128K | Up to 128K |
| API Compatibility | OpenAI ✅ | OpenAI ✅ | OpenAI ✅ | OpenAI ✅ |
The "M" stands for million tokens, by the way, which is the standard unit everyone charges by. If you're new to this, just remember: every chat costs you something, and the difference between $0.01 and $3.50 per million tokens is enormous when you're processing real volumes.
DeepSeek: The One I Reach for Most Often
I'll be real, DeepSeek V4 Flash has become my default model for almost everything. The pricing alone floored me the first time I saw it.
What I Actually Use
- V4 Flash at $0.25 per million output tokens. This is my daily driver.
- V3.2 at $0.38. Newest architecture, but I haven't noticed a huge jump over V4 in everyday prompts.
- V4 Pro at $0.78. For when I need production-quality stuff.
- R1 at $2.50. This is the "reasoner" model — really good at gnarly math problems.
- Coder at $0.25. Specialized for code, but V4 Flash handles code so well I rarely bother switching.
What Made Me a Fan
Honestly, the price-to-quality ratio. I ran V4 Flash through a bunch of coding tasks and it held its own against GPT-4o on everything I tried. It blows my mind that I can get that level of output for a quarter per million tokens. It's also blazing fast — somewhere around 60 tokens per second in my testing, which means no awkward waiting when I'm iterating.
The English is genuinely excellent too. I had no idea a Chinese-built model would feel this natural in English prose.
Where It Frustrates Me
There are a few gaps. First, no native vision capability — you can't throw an image at it and ask what's in it. That's a dealbreaker for some projects. Second, for pure Chinese language tasks, GLM and Kimi edge it out slightly. And third, the model lineup is smaller than Qwen's, so if you need a tiny 1B parameter model for edge deployment, you're out of luck here.
Switching an Existing Project to DeepSeek V4 Flash
This was one of the easiest swaps I've ever done. If you've worked with the OpenAI Python library, you're already 90% of the way there:
from openai import OpenAI
client = OpenAI(
api_key="ga_xxxxxxxxxxxx",
base_url="https://global-apis.com/v1"
)
response = client.chat.completions.create(
model="deepseek-v4-flash",
messages=[{"role": "user", "content": "Explain quantum computing in 100 words"}]
)
print(response.choices[0].message.content)
That's it. The base URL change is the only real difference. I was shocked at how painless it was.
Qwen: The One With a Model for Everything
If DeepSeek is a sharp knife, Qwen is the entire kitchen drawer. Alibaba ships so many variants that keeping track of them feels like a part-time job, but there's almost certainly one that fits whatever you're building.
The Lineup I Keep Coming Back To
- Qwen3-8B at $0.01 per million output tokens. One cent. Yes, really.
- Qwen3-32B at $0.28. The general-purpose pick.
- Qwen3-Coder-30B at $0.35. For when code generation is the bottleneck.
- Qwen3-VL-32B at $0.52. Vision-language, handles images.
- Qwen3-Omni-30B at $0.52. Audio, video, image, all in one.
- Qwen3.5-397B at $2.34. Enterprise-tier reasoning.
I had no idea you could run a vision-capable model for fifty-something cents per million output tokens. That kind of pricing just wasn't part of my mental map before I started this journey.
What I Love
The sheer variety. Whatever size of model you need, whatever modality, there's probably a Qwen for it. The Omni model particularly impresses me — being able to send audio, video, and images through one endpoint simplifies my codebase.
Alibaba's infrastructure is no joke either, so reliability has been solid in my projects.
What Drives Me Crazy
The naming. Honestly. I was poking around and found Qwen3, Qwen3.5, Qwen3.6, and at least three different variants of each. Pick the wrong one and your prompt goes to a totally different model than you expected. Some models also feel overpriced — Qwen3.6-35B at $1 per million output feels steep for what you get.
English quality is fine, but I think DeepSeek edges it slightly on natural-sounding prose.
A Typical Qwen Call
Here's what most of my Qwen scripts look like:
response = client.chat.completions.create(
model="Qwen/Qwen3-32B",
messages=[{"role": "user", "content": "Write a Python function to merge two sorted lists"}]
)
Same client, same library, just swap the model name. The whole Global API setup is gloriously boring in the best way.
Kimi: The Brainy One
Kimi is the model I bring in when I need to think hard. The K2.5 model absolutely shines on logic puzzles, multi-step reasoning, and anything that benefits from careful chain-of-thought.
What I Reach For
Kimi's pricing sits at the premium end — K2.5 runs $3.00 per million output tokens, with their range going up to $3.50. There isn't really a "cheap Kimi" tier, and I was shocked when I first saw those numbers. But when you need it, you need it.
Why It Earns Its Price Tag
In a word: reasoning. K2.5 is the best of these four at math, logic, and problems that require it to hold a long chain of thought together. It also handles Chinese language beautifully — like DeepSeek's English output, I had no idea a model could feel this natural in Mandarin.
The context window goes up to 128K tokens, so you can throw long documents at it without worrying about losing the thread.
Where It Hurts
Speed isn't Kimi's strength. It's noticeably slower than DeepSeek V4 Flash, which makes sense given how much thinking it's doing. There's also no vision model in the lineup, so if you need multimodal, look elsewhere. And the price per token is real — for casual tasks, Kimi feels like using a Ferrari to grab groceries.
GLM: The Budget Champion With Serious Depth
I saved GLM for last because it kept surprising me the more I used it. The GLM-4-9B at $0.01 per million output tokens is the cheapest model in this entire comparison, but the big brother GLM-5 at $1.92 is a heavyweight contender.
The Models Worth Knowing
- GLM-4-9B at $0.01. Tiny. Cheap. Surprisingly capable.
- GLM-5 at $1.92. The flagship, and a true all-rounder.
The price range overall goes from $0.01 to $1.92, which is wild given the spread in capability.
Why I Recommend It
Two things stand out. First, Chinese language handling is best-in-class — Zhipu built this for Chinese tasks and it shows. Second, GLM-4.6V brings vision capabilities into the family, which means you don't have to leave the GLM ecosystem just because a project needs to "see" an image.
For anyone building tools that primarily serve Chinese users, GLM is honestly the strongest pick on this list.
The Tradeoffs
Code generation is weaker than the other three. It's not bad — it's just that DeepSeek and Qwen give you a noticeably better experience for code-heavy work. English language quality is solid but slightly behind DeepSeek in my own side-by-side tests.
How I Actually Use These Together
Here's a piece of advice I wish someone had given me at bootcamp: don't pick one model. Pick the right model for each call.
For my current project, I do something like this:
- Generate boilerplate code? DeepSeek V4 Flash.
- Analyze a user-uploaded image? Qwen3-VL-32B.
- Crunch through a complicated logic problem? Kimi K2.5.
- Handle Chinese-language customer service? GLM-4-9B for cheap routing, GLM-5 for tricky cases.
Splitting traffic this way cut my API bill by about 60% compared to running everything on a single premium model. I was shocked when I saw the actual numbers on my dashboard
Top comments (0)