Four Chinese AI Families, One Month of Testing: My Honest Take

#python #machinelearning #webdev #api

I've been writing open source software for about twelve years now, and somewhere along the way I developed an allergy to walled gardens. You know the ones — proprietary APIs with pricing pages that change overnight, model weights locked behind NDAs, and "developer portals" that feel more like velvet ropes than actual tools. So when Chinese labs started dropping competitive models under Apache 2.0 and MIT licenses, I paid attention. Really paid attention.

For the past month I've been hammering on four major Chinese model families — DeepSeek, Qwen, Kimi, and GLM — through Global API's unified OpenAI-compatible endpoint. I'm not here to sell you on any single one. I'm here to share what actually happened when I put each family through real workloads: code generation, Chinese translation, math reasoning, the boring stuff, and the fun stuff. Bring your favorite terminal, because we're going deep.

Why I Even Bothered

Let me rewind a bit. Two years ago, if you wanted a serious large language model for production work, you basically had two flavors of walled garden: OpenAI or Anthropic. Both great, both proprietary, both with that lovely "we can change the rate limit whenever we want" energy. Self-hosting was a fantasy unless you had eight H100s collecting dust in a closet.

Then the open weights floodgates opened. Meta did its thing with Llama. Mistral shipped Apache-licensed gems. And quietly, in parallel, Chinese labs started releasing weights that genuinely competed — and in some benchmarks beat — the Western incumbents. We're talking Apache 2.0, MIT, custom permissive licenses. The kind of licenses that let you fork, modify, and ship without asking permission from a corporate gatekeeper.

I maintain a few smaller open source projects, and the licensing matters to me. It matters to the companies I consult for who don't want their roadmap hostage to a vendor's quarterly pivot. So I tested all four families with that lens: what can I actually do with this, and what strings are attached?

The Lineup at a Glance

Before I get into the individual breakdowns, here's the high-level map of what I worked with. All four families expose OpenAI-compatible endpoints, which means I could route everything through a single client with a base_url swap. That alone is a beautiful thing — the anti-walled-garden dream realized.

Family	Developer	Price Range (output $/M)	My License Vibe
DeepSeek	DeepSeek (幻方)	$0.25–$2.50	Open weights, custom permissive
Qwen	Alibaba (阿里)	$0.01–$3.20	Apache 2.0 on most releases
Kimi	Moonshot AI (月之暗面)	$3.00–$3.50	More closed, weights restricted
GLM	Zhipu AI (智谱)	$0.01–$1.92	MIT on the smaller models, custom for flagship

Context windows topped out at 128K across the board. All four speak the OpenAI chat completions dialect, so the switching cost between them was basically zero. That last part deserves a moment of appreciation, because it used to be a nightmare.

DeepSeek: The Speed Demon I Keep Coming Back To

I'll be honest: DeepSeek V4 Flash became my daily driver within about three days of testing. At $0.25 per million output tokens, it does things that would have cost me real money a year ago. I ran my usual battery of HumanEval-style coding problems, some MBPP-flavored Python exercises, and a handful of ad-hoc "explain this regex to me" requests. V4 Flash handled them all without making me wait, spitting out tokens at roughly 60 tokens per second on my test runs.

The whole DeepSeek stack feels like it was built by people who actually use these things. Here's the lineup I leaned on:

Model	Output $/M	What I Used It For
V4 Flash	$0.25	Default everything — coding, summaries, drafting
V3.2	$0.38	When I wanted the newer architecture for tricky refactors
V4 Pro	$0.78	Production workloads where I needed extra polish
R1 (Reasoner)	$2.50	Math proofs, logic puzzles, anything with steps
Coder	$0.25	Specialized code-completion loops

The code generation is genuinely top-tier. I'm not exaggerating when I say V4 Flash produced output that I would have assumed came from GPT-4o twelve months ago. English handling is excellent, which surprised me — I'd been subtly biased toward assuming Western models would always win on English prose. Nope. DeepSeek holds its own.

Where it stumbles: vision is limited. There's no native image understanding path on V4 Flash or V4 Pro that I could find, which is a real gap if your workflow involves screenshots or diagrams. Chinese-language output is solid but not class-leading — both Kimi and GLM edged it out on translation quality in my side-by-side checks. And the model variety is narrower than what Qwen offers; you're choosing between maybe five serious options instead of fifteen.

But here's the thing that sealed it for me: the open-weight heritage. DeepSeek publishes research, releases weights, and generally behaves like a lab that wants you to understand what's happening under the hood. That's not nothing. That's the entire vibe I want from my AI tooling.

A typical V4 Flash call looks like this for me:

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": "Refactor this Python function to use generators instead of building a list."}
    ]
)
print(response.choices[0].message.content)

Clean, boring, works. The way I like my APIs.

Qwen: The Swiss Army Knife With Apache on the Side

Alibaba's Qwen family is the one I describe to friends as "the buffet." Whatever weird size or modality you're hunting for, there's probably a Qwen variant. From the absolutely-tiny Qwen3-8B at $0.01 per million output tokens (yes, a penny per million) all the way up to the enterprise-grade Qwen3.5-397B at $2.34, the range is wild.

Here's what I had on rotation:

Model	Output $/M	My Use Case
Qwen3-8B	$0.01	Classification, tiny tweaks, anything where speed beats nuance
Qwen3-32B	$0.28	The generalist — my second-most-used model behind V4 Flash
Qwen3-Coder-30B	$0.35	When I needed more code-specific tuning
Qwen3-VL-32B	$0.52	Image understanding tasks (yes, it actually works)
Qwen3-Omni-30B	$0.52	Audio + video + image in one shot
Qwen3.5-397B	$2.34	The big guns for gnarly reasoning chains

The multimodal story is where Qwen genuinely pulls ahead. Qwen3-VL handled my screenshot-to-markdown tests without embarrassing itself, and Qwen3-Omni's audio transcription was good enough that I'm using it in a personal project right now. If your work touches anything beyond pure text, Qwen deserves a serious look.

The Apache 2.0 licensing on most of the smaller models is the cherry on top. I can take Qwen3-8B, fine-tune it on my own data, ship the resulting weights, and never once ask Alibaba for permission. That's the open source ethos I want to defend, and Qwen delivers.

Complaints? The naming is a nightmare. Qwen3, Qwen3.5, Qwen3.6, with -VL and -Omni suffixes flying around — I had to keep a cheat sheet pinned to my monitor. Mid-range English is good but not DeepSeek-tier in my testing. And a couple of the mid-tier models feel slightly overpriced for what they deliver. Qwen3.6-35B at roughly a dollar per million tokens is the kind of pricing that makes me raise an eyebrow when V4 Flash exists for a quarter of that.

A typical Qwen call for general-purpose work:

response = client.chat.completions.create(
    model="Qwen/Qwen3-32B",
    messages=[
        {"role": "user", "content": "Write a Python function to merge two sorted lists without using sorted()."}
    ]
)
print(response.choices[0].message.content)

Same client, same base_url, different model name. This is the dream. No separate SDK, no vendor-specific quirks, no "actually for this one feature you need to use our custom Python wrapper." Just OpenAI-compatible and done.

Kimi: The Reasoning Outlier

Now here's where my open source heart got a little conflicted. Kimi from Moonshot AI put up the best reasoning benchmark numbers of the four families, full stop. When I threw genuinely hard problems at it — the kind of multi-step logic puzzles where you'd normally break out a notebook — K2.5 consistently outperformed the field. There's a reason it has the reputation it does.

The pricing, though. Kimi lives in a different neighborhood:

Model	Output $/M	Notes
K2.5	$3.00	The flagship reasoning model

Kimi's pricing band sits at $3.00–$3.50 per million output tokens, which is roughly an order of magnitude above DeepSeek's budget tier. For casual workloads, that adds up fast. For workloads where reasoning quality is the whole point — theorem proving, complex code synthesis, multi-hop research — the premium is justifiable.

But here's my hangup: Kimi is the most closed of the four families. The flagship weights aren't freely available in the way DeepSeek's or Qwen's are. You can call the API, you can pay the bill, you cannot self-host K2.5 on your own hardware. That's a dealbreaker for some of the projects I care about, and I suspect it is for yours too if you've read this far.

Kimi also lacks a vision/multimodal path in my testing. Text in, text out, no images, no audio. If your pipeline demands multimodal, look elsewhere.

For pure reasoning tasks where budget isn't the primary constraint, Kimi earns its reputation. I just wish the licensing story were better. A $3/M model that I can't inspect, can't fine-tune, can't run on my own metal — that's a vendor relationship, not an open source tool.

GLM: The Quiet MIT-Licensed Champion

If you've been following Chinese AI releases at all, you know Zhipu AI has been shipping GLM models at a relentless pace. What you might not know is how good the licensing has gotten. GLM-4-9B ships under MIT. Yes, MIT. The most permissive license in mainstream use, the one that basically says "do whatever you want, just keep the copyright notice."

Here's the GLM lineup I tested:

Model	Output $/M	Sweet Spot
GLM-4-9B	$0.01	Penny pricing for tiny tasks
GLM-5	$1.92	Flagship, GLM-4.6V handles vision

GLM-4-9B at a penny per million output tokens is, as far as I can tell, the cheapest serious LLM API on the planet right now. I used it for routing decisions, lightweight classification, and as a "first pass" filter before escalating to a bigger model. The total bill for a week of constant background use was, and I'm not making this up, less than a fancy coffee.

GLM-5 at $1.92 per million is the flagship, and it's no slouch. Chinese-language performance is the best of the four families — no surprise, given Zhipu's roots — and English quality holds its own against the mid-tier competition. GLM-4.6V brings vision into the mix, which rounds out the multimodal story that DeepSeek and Kimi are missing.

The price band sits at $0.01–$1.92 per million, with the flagship undercutting Kimi's cheapest model by a comfortable margin. For a developer who wants MIT-licensed weights they can actually ship, GLM is the standout.

If I had to pick one family for an open source project that needed to be defensible five years from now, it might be GLM. The combination of permissive licensing, strong Chinese and English performance, vision support, and reasonable pricing is hard to beat.

What I Actually Deployed

After a month of testing, here's where I landed for my own projects:

V4 Flash for general coding, content drafting, and anything where I wanted the best token-per-dollar ratio. My default.
Qwen3-32B as the backup generalist. Switched to it when V4 Flash had a momentary hiccup or when I wanted a second opinion.
GLM-4-9B as the routing/classification layer. A penny per million means I can call it a thousand times without thinking.
GLM-5 for Chinese-heavy work where I needed the extra quality.
Kimi K2.5 for the rare, genuinely hard reasoning task. Reserved for when nothing else cut it.
Qwen3-VL-32B when I needed to process an image.