DEV Community

loyaldash
loyaldash

Posted on

I Ran Chinese AI Models Through Real Client Work: Here's What Stuck

I Ran Chinese AI Models Through Real Client Work: Here's What Stuck

Look, I'm not a tech journalist with a corporate expense account. I'm a freelance dev who bills by the hour, and every API call comes out of my pocket before the client ever sees an invoice. So when I started hearing about the Chinese model scene getting genuinely competitive, I did what any 精打细算 freelancer would do: I opened a spreadsheet, signed up for Global API, and started running real client workloads through four different model families to see which ones actually deserve a line item on my invoice.

I tested DeepSeek, Qwen, Kimi, and GLM over the course of about a month. Real deliverables — client blog posts, code refactors, that one weird Chinese-language localization project that landed in my lap last quarter. This isn't a benchmark theater piece. This is what happens when someone running a side hustle has to justify every tenth of a cent.

Let me walk you through what I found.


The Lineup and Why I Bothered

The Chinese AI ecosystem has been on my radar since DeepSeek's open-weight models started making noise. Then Qwen kept shipping new versions faster than I could keep track. Kimi kept showing up on reasoning benchmarks. And GLM — well, I had a Mandarin client, so GLM was non-negotiable anyway.

The challenge wasn't whether to use them. It was figuring out which one to reach for when the timer is running on a billable hour.

All four model families offer OpenAI-compatible APIs, which means I could route everything through a single endpoint at https://global-apis.com/v1 and compare apples to apples without rewriting my client code four times. That alone saved me hours of integration work — and hours are money, remember.


Kimi: The Brainy One I Couldn't Afford to Love

I'll be honest — I wanted Kimi to be my daily driver. The reasoning quality on K2.5 at $3.00/M output is genuinely impressive. I threw some architecture questions at it (the kind I'd normally ask Claude for), and the answers were sharp. Thoughtful. Actually useful for the kind of design docs I bill for.

But here's the math problem: $3.00 per million output tokens. Let me translate that into freelancer-speak. If I'm generating a 2,000-word client deliverable — say, a technical specification document — I'm burning through roughly 2,500 output tokens. That's about $0.0075 per document. Sounds tiny, right?

Now multiply by 30 documents a month. Now add the follow-up questions, the revisions, the "can you expand on section three" emails. By the end of testing, my Kimi bill was eating about $18-22 a week for moderate use. That's $80+ a month just for one model on one client engagement. If I'm billing that client $75/hour, I need to charge them for the AI time OR eat the cost. Neither feels great when my margins are already thin.

Strengths worth mentioning:

  • Reasoning benchmarks are top-tier — genuinely best-in-class for complex logic
  • Output quality feels polished, less editing required on my end
  • Handles long context beautifully (up to 128K)

Weaknesses that hurt my wallet:

  • No budget option — everything is $3.00-$3.50/M, which is premium pricing
  • Slower than the alternatives, which means I'm waiting around while billing
  • No vision/multimodal support

For deep thinking tasks where quality matters more than cost? Kimi earns its keep. For day-to-day client work? Hard to justify at these prices.


GLM: The Mandarin Workhorse

One of my clients is a Shanghai-based e-commerce startup. Their content needs to sound natural in Simplified Chinese — not "translated from English" natural, but actual 原生 native-feeling copy. This is where GLM shines, and it's not even close.

GLM-4-9B at $0.01/M output is basically free. I use it for first drafts of Chinese product descriptions, then spend my billable hours on the polishing pass instead of the blank-page problem. The cost difference versus hiring a human copywriter for first drafts is laughable.

For bigger jobs, GLM-5 at $1.92/M is my go-to. The quality jump is noticeable — better nuance, fewer of those weird literal-translation artifacts that scream "AI wrote this." When the deliverable matters and the client is paying premium rates for native-quality content, GLM-5 earns its price tag.

The numbers that matter to me:

  • GLM-4-9B at $0.01/M — I ran probably 50,000 tokens through this last month and spent less than a dollar
  • GLM-5 at $1.92/M — heavier lifting, but still cheaper than the Western premium tier
  • GLM-4.6V handles image tasks when I need OCR or visual analysis on Chinese marketing materials

Weaknesses:

  • English output is good but not best-in-class — I'd still pick DeepSeek for English-heavy work
  • Smaller model ecosystem compared to Qwen
  • Speed is decent but not blazing

For Chinese-language client work specifically, GLM is the obvious call. The $0.01 entry point means I can use it as throwaway infrastructure without thinking twice.


DeepSeek: The Surprise Winner of My Spreadsheet

Here's where things got interesting. I went into this thinking DeepSeek would be "the budget option" — you know, good enough, cheap, nothing fancy. I was wrong.

V4 Flash at $0.25/M output is the model that changed how I think about API costs. Let me do the math on what that means for my actual business:

A typical code generation task — say, writing a Python function with documentation — burns maybe 400 output tokens. At $0.25/M, that's $0.0001 per generation. I could run 10,000 of those for a dollar. During a heavy refactoring sprint, I might do 200 generations in a day. We're talking about $0.02 in API costs. That's less than the electricity to run my monitor.

I started routing my English-language coding work through DeepSeek V4 Flash. Same OpenAI-compatible interface, so the migration took about four minutes. The quality held up against much pricier models on HumanEval-style tasks. For the bulk of client work — boilerplate generation, test writing, documentation, straightforward refactors — there's no reason to pay more.

The other models in the DeepSeek family I tested:

  • V3.2 at $0.38/M — the latest architecture, slightly more capable
  • V4 Pro at $0.78/M — when I need production-grade output quality
  • R1 (Reasoner) at $2.50/M — for genuinely hard logic problems, not a daily driver
  • Coder at $0.25/M — code-specific, which I mostly skipped since V4 Flash already handles code well

The thing about DeepSeek that impressed me most wasn't any single benchmark. It was the consistency. I never had a "wait, what is this output" moment. For client deliverables, predictability is worth more than occasional brilliance.

Weaknesses:

  • Limited vision capabilities — no native image understanding
  • Chinese output is good but trails GLM and Kimi on nuanced tasks
  • Fewer model size options than Qwen offers

But honestly? For a freelancer doing English-language work, these weaknesses barely register.


Qwen: The Swiss Army Knife I Keep Coming Back To

If I had to pick one model family to standardize on, it'd be Qwen. Not because it's the best at any one thing, but because it has a model for every budget tier and every task type I've encountered.

The price spread is wild: $0.01/M for Qwen3-8B all the way up to $2.34/M for Qwen3.5-397B. That means I can use the same API endpoint for everything from "summarize this email" to "reason through this enterprise architecture question."

My actual Qwen usage breaks down like this:

  • Qwen3-8B at $0.01/M for lightweight stuff: classification, simple extractions, short summaries. This is so cheap I forget I'm paying for it.
  • Qwen3-32B at $0.28/M for general-purpose work. This is probably my most-used model across the whole testing period. The quality-to-price ratio is excellent.
  • Qwen3-Coder-30B at $0.35/M when I need code generation that's a step above V4 Flash. Honestly? DeepSeek's V4 Flash is usually good enough, but Qwen3-Coder handles some edge cases better.
  • Qwen3-VL-32B at $0.52/M when a client sends me a screenshot or mockup and I need to extract requirements. Vision capability is genuinely useful for client work — half my intake is via screenshots.
  • Qwen3-Omni-30B at $0.52/M — the multimodal option. I haven't used this as much as I expected, but when I needed it (audio transcription for a podcast client), it worked fine.

The Alibaba backing means this isn't going anywhere. The infrastructure is enterprise-grade, and they're shipping new models constantly. When I'm committing client work to a model family, I want to know the vendor isn't going to vanish in six months.

Weaknesses:

  • The naming is genuinely confusing. Qwen3-8B, Qwen3-32B, Qwen3.5-397B, Qwen3-VL-32B... I keep a cheat sheet open. There's no logical progression in the version numbers.
  • English output is good but not DeepSeek-level on certain nuanced tasks
  • Qwen3.6-35B at $1/M feels overpriced for what it delivers — I'd rather pay $0.28 for Qwen3-32B

But the range of options means there's almost always a Qwen model that fits the budget for any given task.


What I Actually Spent: The Honest Numbers

Time for the spreadsheet reveal. Across about four weeks of real client work mixed with testing:

  • Kimi: roughly $85 (premium reasoning tasks only, rationed carefully)
  • GLM: about $12 (mostly GLM-4-9B at $0.01/M, with some GLM-5)
  • DeepSeek: around $8 (heavy usage on V4 Flash for coding)
  • Qwen: maybe $15 (broad usage across multiple models)

Total: about $120 for a month of meaningful AI-assisted client work. That's not nothing, but it's way less than the equivalent in pure Western premium model spend would have been. And the productivity boost — having a competent AI assistant available at marginal cost — easily justifies it.

The real win is having the cheap options. Qwen3-8B at $0.01/M means I can run a classifier or summarizer over a huge batch of client documents without flinching. That's a category of work that simply wasn't viable at premium pricing.


My Actual Code Setup

Since I route everything through Global API's unified endpoint, my setup is stupidly simple. Here's how I typically call DeepSeek V4 Flash for a client coding task:

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": "system", "content": "You are a senior Python developer. Write clean, production-ready code."},
        {"role": "user", "content": "Write a function to validate email addresses using regex, with full type hints and docstring."}
    ],
    temperature=0.2
)

print(response.choices[0].message.content)
Enter fullscreen mode Exit fullscreen mode

Switching to Qwen3-32B for general-purpose work takes about two seconds:

response = client.chat.completions.create(
    model="Qwen/Qwen3-32B",
    messages=[
        {"role": "user", "content": "Summarize the key points from this client meeting transcript: [paste transcript here]"}
    ],
    temperature=0.5
)
Enter fullscreen mode Exit fullscreen mode

Same client object, same auth, same endpoint. The only thing that changes is the model name. This is huge for billable work — I'm not maintaining four different API integrations, debugging four different SDK quirks, or managing four different rate limit policies.


Which One Should You Actually Use?

If you're a freelancer running a side hustle and watching every dollar:

  • Default to DeepSeek V4 Flash for English coding and content work. The $0.25/M output is the sweet spot.
  • Use Qwen3-8B ($0.01/M) for batch tasks where you need volume. Classification, extraction, bulk summarization.
  • Reach for Qwen3-32B ($0.28/M) when you need general-purpose quality that's a step above the budget tier.
  • Pay for Kimi K2.5 ($3.00/M) only when the reasoning quality genuinely matters — complex architecture decisions, hard debugging, math-heavy work.
  • Use GLM-4-9B ($0.01/M) and GLM-5 ($1.92/M) when Chinese-language quality is

Top comments (0)