I Spent Two Weeks Testing Chinese AI Models and Got Surprised

I Spent Two Weeks Testing Chinese AI Models and Got Surprised

Let me be real with you. Before I started digging into this stuff, I was the kind of bootcamp grad who just defaulted to GPT-4o for everything. That's what my instructors used, that's what my classmates used, so I never questioned it. Then I accidentally stumbled onto a thread about Chinese AI models and my brain kind of broke.

What followed was two weeks of testing, comparing pricing pages, and genuinely geeking out. I figured I should write this up because I wish someone had handed me a simple breakdown before I spent my own money figuring this out.

How I Even Got Into This Rabbit Hole

It started with a Slack message. Someone posted a link saying DeepSeek could do basically the same thing as GPT-4o for like 97% less money. I clicked expecting a clickbait article and instead fell down a hole I haven't climbed out of yet.

I had no idea China had built this many serious AI models. DeepSeek, Qwen, Kimi, GLM. Four whole families of models, each with multiple versions, each with different prices. I genuinely thought the AI world was just OpenAI, Anthropic, and Google calling the shots.

Spoiler: it's way more crowded than I realised.

The TL;DR from my testing: DeepSeek V4 Flash is the winner for most people on a budget. Qwen has more model options than any other family. Kimi is a monster at reasoning tasks. GLM absolutely crushes Chinese language stuff.

The Pricing Reality That Blew My Mind

I'm just gonna paste the pricing ranges here because this is the part that shocked me the most:

• DeepSeek: $0.25–$2.50 per million output tokens
• Qwen: $0.01–$3.20 per million output tokens
• Kimi: $3.00–$3.50 per million output tokens
• GLM: $0.01–$1.92 per million output tokens

I had to triple-check these numbers. A model from Qwen costs literally one cent per million tokens. ONE CENT. I was paying GPT-4o prices like a chump.

For reference, one million tokens is roughly 750,000 words. So if I'm paying $0.01 per million, I could process an entire novel for less than the cost of a gumball.

Let me give you the specific models I tested and what they cost:

DeepSeek Models

• V4 Flash: $0.25/M (my daily driver now)
• V3.2: $0.38/M
• V4 Pro: $0.78/M
• R1 (Reasoner): $2.50/M
• Coder: $0.25/M

Qwen Models

• Qwen3-8B: $0.01/M (yes, really)
• Qwen3-32B: $0.28/M
• Qwen3-Coder-30B: $0.35/M
• Qwen3-VL-32B: $0.52/M (vision)
• Qwen3-Omni-30B: $0.52/M (handles audio, video, image)
• Qwen3.5-397B: $2.34/M
• Qwen3.6-35B: $1.00/M

Kimi Models

• K2.5: $3.00/M (their flagship)

GLM Models

• GLM-4-9B: $0.01/M
• GLM-5: $1.92/M
• GLM-4.6V: handles vision tasks

My Honest Take on Each Family

DeepSeek: The One I Keep Coming Back To

Okay so DeepSeek is the one that genuinely changed how I work. Their V4 Flash model at $0.25 per million tokens does basically everything I used to use GPT-4o for. I threw some coding problems at it, content writing prompts, random brainstorming stuff, and it kept delivering.

The thing that really surprised me was the code generation. I'm not great at algorithms yet (bootcamp only got me so far), and DeepSeek's code output has been consistently top-tier. There's a benchmark called HumanEval that tests code generation quality, and DeepSeek is right at the top of the leaderboard for models in this price range.

They also have a dedicated Coder model that costs the same as V4 Flash ($0.25/M) if you want something specifically tuned for programming tasks. I tested it on some LeetCode-style problems and it held its own.

The downsides? DeepSeek doesn't really do vision stuff well. If you need to analyze images, you're gonna want a different model. Their Chinese language skills are also a hair behind GLM and Kimi, though they're still solid. And they have fewer model size options compared to Qwen.

Speed-wise, I was getting around 60 tokens per second from V4 Flash, which is one of the fastest models I've tested. That's roughly the speed of reading a short paragraph every second, which feels instant when you're chatting with it.

Here's the basic Python code I used to test DeepSeek via Global API:

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)

The base URL trick is what makes this work. Global API gives you one endpoint and you swap model names to switch between providers. It's wild.

Qwen: The Family That Has Literally Everything

Qwen (made by Alibaba, which I didn't realise until this whole journey) is like the Swiss Army knife of Chinese AI. Whatever weird thing you want to do, there's probably a Qwen model for it.

Need image understanding? Qwen3-VL-32B at $0.52/M. Want audio and video together? Qwen3-Omni-30B at $0.52/M. Need to save money on basic stuff? Qwen3-8B at $0.01/M. Got a giant enterprise workload? Qwen3.5-397B at $2.34/M. There's like ten different models in this family.

The pricing spread is genuinely insane. You can go from $0.01 to $3.20 per million tokens just within the Qwen family. I think this is because Alibaba releases so many versions and they price each one based on what it does.

My favorite in the bunch was Qwen3-32B at $0.28/M. It hits this sweet spot where it's cheap enough to use casually but smart enough to handle most tasks. For general coding and content work, it performed better than I expected.

One thing I found mildly annoying: the naming. Qwen3.5, Qwen3.6, Qwen3-32B, Qwen3-Omni-30B. After my third day of testing I was mixing them up constantly. Maybe that's just me being a beginner though.

Here's how I tested Qwen3-32B:

response = client.chat.completions.create(
    model="Qwen/Qwen3-32B",
    messages=[{"role": "user", "content": "Write a Python function to merge two sorted lists"}]
)
print(response.choices[0].message.content)

Notice I'm using the same client object? That's the beauty of Global API. One setup, swap the model name, done.

Kimi: The Brainy One

Kimi is made by Moonshot AI, which honestly sounds like a sci-fi company name. Their flagship is K2.5 at $3.00/M, which puts them firmly in the premium tier. They're not trying to compete on price, they're competing on being the smartest.

And honestly? It kinda worked. When I threw reasoning-heavy problems at K2.5, it outperformed everything else I tested. Math problems, logic puzzles, multi-step reasoning, that kind of stuff. Kimi is the one I'd pick if I was building something that needed to actually think, not just generate text.

The catch is the price. At $3.00/M for K2.5, you're paying GPT-4o-level money. So you'd want to use it strategically for the tasks that actually need its reasoning power, not for casual chitchat.

I don't have as much to say about Kimi because they have fewer model options and they're all premium-priced. But if reasoning is your bottleneck, it's worth a serious look.

GLM: The Chinese Language Champion

GLM comes from Zhipu AI. They're the specialists when it comes to Chinese language tasks. If you're building anything for a Chinese-speaking audience, GLM should probably be your first stop.

Their model lineup:

• GLM-4-9B: $0.01/M (ultra-budget)
• GLM-5: $1.92/M (flagship)
• GLM-4.6V: vision capabilities

GLM-4-9B at one cent per million tokens is honestly absurd. For simple Chinese text tasks, that's basically free. GLM-5 at $1.92/M is their heavy hitter and handles more complex work.

For English tasks, GLM is good but not exceptional. It's like a B+ student. Does the work, gets it done, but isn't gonna blow your mind. The real magic happens when you feed it Chinese prompts.

They also have GLM-4.6V which handles images, which is nice because DeepSeek doesn't really do vision well.

How I Actually Use All Four Now

After two weeks, here's what I landed on for my personal workflow:

For everyday coding questions and content writing, DeepSeek V4 Flash is my go-to. The $0.25/M price means I don't even think about how much I'm using it.

When I need vision stuff (analyzing screenshots, reading images), I use Qwen3-VL-32B. It's $0.52/M which is reasonable for vision tasks.

If I'm working on something that requires heavy reasoning, like complex algorithms or math, I switch to Kimi K2.5. Yes it's $3.00/M but for specific hard problems it's worth it.

For any Chinese language work (which is rare for me but happens), I use GLM.

The fact that I can swap between all four using the same API endpoint is what made this actually practical for me as a beginner. I didn't have to learn four different SDKs or sign up for four different accounts. Just one base URL, one API key, and I change the model name.

What Surprised Me the Most

Honestly? A few things:

1. The price gap is real. I was spending maybe $40/month on GPT-4o for personal projects. After switching to DeepSeek for daily stuff, my last bill was under $5.

2. The quality gap is smaller than I expected. For most tasks I'd argue DeepSeek V4 Flash is within 5-10% of GPT-4o for a fraction of the cost.

3. The Chinese AI ecosystem is huge. Four major model families, dozens of variants, all competing seriously. This isn't some second-tier alternative market.

4. OpenAI-compatible APIs make everything easier. Every single one of these providers uses the same API format as OpenAI, which means switching is mostly just changing a model name.

The Honest Caveats

I'm a bootcamp grad. I'm not a researcher. I'm not running massive enterprise workloads. My testing was personal projects, some work for clients, and a lot of just messing around.

If you're building production systems handling millions of requests, you should probably do your own more rigorous testing. The benchmarks I mentioned (HumanEval, MBPP) are what the industry uses, but real-world performance varies.

Also, the pricing I quoted is what I saw at the time of testing. These models and prices update regularly. Qwen especially seems to release new versions every few months.

Try It Yourself

If any of this sounds interesting, I found the easiest way to test all these models is through Global API. They give you one endpoint and access to all four families I talked about, plus a bunch of other models. You don't need separate accounts for each provider.

I went to global-apis.com, signed up, got an API key, and was making my first DeepSeek request within like ten minutes. The setup with the OpenAI Python library is exactly what I showed in the code examples above.

Whether you're a bootcamp grad like me or you've been doing this longer, it's worth poking around. The pricing alone will probably make you do a double-take.

Just check it out if you're curious. That's what I'd say. No pressure, but seriously, once you see the price difference, it's hard to unsee it.