DEV Community

RileyKim
RileyKim

Posted on

I Tested DeepSeek vs Qwen vs Kimi vs GLM for a Month — Here's What Won

I Tested DeepSeek vs Qwen vs Kimi vs GLM for a Month — Here's What Won

ok so heres the thing. I've been building AI-powered tools for like 3 years now, and honestly, I got tired of paying OpenAI prices. Like REALLY tired. My last invoice made me want to cry, no joke.

So I did what any reasonable indie hacker would do — I went down the Chinese AI rabbit hole. DeepSeek, Qwen, Kimi, GLM. All the big ones. I ran them through their paces for a solid month on real client work, and honestly? Some of these are WILD good for the price. Like, embarrassingly good.

This is what I found.


The TL;DR Before I Ramble

Heres the short version before I get into the weeds:

  • DeepSeek V4 Flash = best bang for your buck, period
  • Qwen = most options, basically a model for everything
  • Kimi = smartest of the bunch, but you pay for it
  • GLM = the secret weapon if you're doing Chinese language stuff

I tested everything through Global API's unified endpoint because honestly, I cannot be bothered to manage 4 different API keys and dashboards. More on that at the end.


What I Actually Tested

Before I get into opinions, heres what I threw at these models:

  • Code generation (Python, JS, some Rust when I was feeling brave)
  • Long-form content writing
  • Translation (EN <-> ZH mostly)
  • Reasoning tasks (math, logic puzzles, the usual)
  • Image understanding on the models that support it
  • Speed benchmarks (tokens/sec, latency)

I ran the same prompts across all four families, multiple times, and tracked output quality + cost.


The Numbers (Real Prices, No BS)

Model Family Price Range (output $/M) Budget Pick Flagship
DeepSeek $0.25 – $2.50 V4 Flash @ $0.25 V4 Pro @ $0.78
Qwen $0.01 – $3.20 Qwen3-8B @ $0.01 Qwen3.5-397B @ $2.34
Kimi $3.00 – $3.50 K2.5 @ $3.00 K2.5 @ $3.00
GLM $0.01 – $1.92 GLM-4-9B @ $0.01 GLM-5 @ $1.92

Yeah, you read that right. Qwen3-8B is ONE CENT per million output tokens. ONE CENT. I've spent more on a vending machine.


DeepSeek: My Daily Driver Now

honestly, I gotta say, DeepSeek kinda ruined OpenAI for me. V4 Flash is my go-to for like 80% of tasks now.

What I'm Actually Using

V4 Flash ($0.25/M output) — this is the workhorse. I use it for:

  • Writing product descriptions
  • Generating code snippets
  • Summarizing documents
  • Pretty much everything boring

V3.2 ($0.38/M) — slightly newer architecture, I tested it but honestly didn't see enough diff to switch from Flash for my use cases.

V4 Pro ($0.78/M) — when I need production-grade output and can't afford mistakes.

R1 ($2.50/M) — the reasoner. Use this when math or logic is involved. Its expensive but it can actually THINK through problems.

Coder ($0.25/M) — same price as Flash but fine-tuned for code. Honestly, I found regular V4 Flash to be just as good for code, so I didn't end up using this much.

What I Like

  • The price-to-performance is INSANE. Like, V4 Flash at $0.25/M genuinely rivals GPT-4o on most of my tasks
  • Speed is no joke — I was getting around 60 tokens/sec, which means responses stream fast
  • English output is clean, no weird artifacts
  • Code generation is honestly top-tier. Its been killing it on HumanEval-style stuff for me

What Bugged Me

  • Vision is basically non-existent. If I need to look at images, I have to use something else
  • Chinese language performance is good but not the BEST (we'll get to that)
  • Not as many model size options as Qwen

How I Use It

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)
Enter fullscreen mode Exit fullscreen mode

literally just point the OpenAI SDK at Global API's endpoint and you good. Took me like 2 minutes to switch over.


Qwen: The "I Need a Model for That Too" Family

Qwen is wild because Alibaba basically released a model for EVERY use case. Its like the Swiss Army knife of Chinese AI.

The Lineup

Qwen3-8B ($0.01/M) — ultra cheap. Use for simple stuff, classification, basic Q&A. Honestly I forget this exists because I just use V4 Flash for everything, but for HIGH volume low-stakes stuff? $0.01 is unbeatable.

Qwen3-32B ($0.28/M) — my second favorite Qwen model. Great general purpose, basically DeepSeek V4 Flash competitor but slightly different vibes.

Qwen3-Coder-30B ($0.35/M) — code specialist. I tested it, it was solid, but again — V4 Flash handled my code needs fine.

Qwen3-VL-32B ($0.52/M) — vision-language model. This is where Qwen pulls ahead. If you need to look at images, this is one of your only options in this comparison.

Qwen3-Omni-30B ($0.52/M) — multimodal. Audio, video, image, text all in one. I haven't really stress-tested this because I don't have a use case but its cool that it exists.

Qwen3.5-397B ($2.34/M) — the big boy. Enterprise reasoning. 397 BILLION parameters. I tested it and yeah, its smart. But its $2.34/M and honestly for what I do, not worth it over cheaper options.

What I Like

  • The RANGE. From $0.01 to $3.20, theres a model for literally every budget
  • Vision models are solid (Qwen3-VL series)
  • Omni-modal is genuinely cool, audio + video + image in one model
  • Alibaba backs it so the infra is rock solid

What Annoyed Me

  • The NAMING. Qwen3, Qwen3.5, Qwen3.6, Qwen3-VL, Qwen3-Omni — I literally had to make a spreadsheet to track whats what
  • English is good but not DeepSeek-level for my taste
  • Some models feel overpriced. Qwen3.6-35B at $1/M for what you get? No thanks.

My Code Example for Qwen

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)
Enter fullscreen mode Exit fullscreen mode

Same client setup, just swap the model name. easy peasy.


Kimi: The Brainy One (That Costs a Lot)

ok Kimi is the one that impressed me the most AND frustrated me the most. Let me explain.

Whats Available

K2.5 ($3.00/M output) — this is the main one, the flagship. Its also the cheapest Kimi model. Yeah, you read that right — Kimi starts at $3.00/M. Which is... a lot compared to the others.

The price range goes up to $3.50/M for their bigger models, but honestly K2.5 is the one I kept coming back to because at least its the "cheapest" option.

What I Like (And I Genuinely Do)

  • The REASONING. Holy cow. I threw some logic puzzles at it that made other models fail, and K2.5 just... got it right
  • Math performance was the best of the bunch
  • For hard problems where you actually need the model to THINK, Kimi is the winner
  • Output quality on complex tasks is noticeably better than the cheaper models

What Made Me Sad

  • The PRICE. $3.00/M is rough when DeepSeek V4 Flash is $0.25/M. That's 12x more expensive
  • Speed is the slowest of the four. Responses take longer to start streaming
  • No vision support. NONE. If you need image stuff, look elsewhere
  • For 90% of my actual use cases, V4 Flash or Qwen3-32B is "good enough" and I save a TON of money

When I'd Actually Use Kimi

Honestly? Only when I hit a wall with the cheaper models. Like, if I'm working on a hard algorithm problem or a complex reasoning task, I'll route it to Kimi. But for general use? No way I'm paying 12x for marginal quality improvements.


GLM: The Underrated One (Especially for Chinese)

I slept on GLM at first. I was like "eh, another Chinese model" and moved on. But then I tested it properly and... damn.

The Models

GLM-4-9B ($0.01/M) — ultra cheap, similar tier to Qwen3-8B. Good for simple tasks.

GLM-5 ($1.92/M) — the flagship. $1.92/M puts it in the "premium but not crazy" range. For context, thats about 7-8x V4 Flash but way cheaper than Kimi.

GLM-4.6V — the vision model. This is actually pretty solid for image understanding.

What Genuinely Surprised Me

  • CHINESE LANGUAGE. Holy crap. GLM is Zhipu AI's baby, and it shows. Chinese output quality is the BEST of the four. Like, noticeably better
  • GLM-5 at $1.92/M is a sweet spot for premium quality without going bankrupt
  • The 9B model at $0.01/M is great for high-volume Chinese tasks
  • Vision capabilities are good (GLM-4.6V)

What I Didn't Love

  • English is good but a tier below DeepSeek for my taste
  • Speed is fine but not blazing
  • Smaller model ecosystem than Qwen
  • Less name recognition means less community examples to learn from

Who Should Care

If you're building anything for Chinese users, GLM should be at the TOP of your list. For pure English work, DeepSeek still wins for me. But the moment Chinese comes into play, GLM pulls ahead.


The Showdown: What I Actually Use Day-to-Day

After a month of testing, heres what my actual setup looks like:

70% of requests → DeepSeek V4 Flash ($0.25/M)

  • Content writing, code, general Q&A
  • Best quality-per-dollar I can find

20% of requests → Qwen3-32B ($0.28/M)

  • When I want a second opinion
  • General purpose fallback

8% of requests → GLM-5 ($1.92/M)

  • When quality really matters
  • Any Chinese language work

2% of requests → Kimi K2.5 ($3.00/M)

  • Hard reasoning tasks
  • When cheaper models fail and I NEED the right answer

My monthly bill went from like $400 on OpenAI to around $60 now. SIXTY DOLLARS. For roughly the same output quality. Honestly, I kinda wish I'd done this sooner.


Speed Rankings (From My Tests)

  1. DeepSeek V4 Flash — ~60 tok/sec, the fastest
  2. Qwen3-32B — close second, smooth streaming
  3. GLM-5 — middle of the pack
  4. Kimi K2.5 — noticeably slower, but you get what you pay for (better reasoning)

If you need instant responses for a chat UI, DeepSeek is your pick. If you can wait a beat for smarter answers, Kimi works.


What I Wish Someone Told Me Before I Started

  • Don't sleep on the cheap models. Qwen3-8B and GLM-4-9B at $0.01/M are insanely good for what they cost
  • Routing is the move. Don't just pick one model, route requests based on task complexity
  • Test on YOUR data. My use case is different from yours. I literally A/B tested for a month before committing
  • The OpenAI-compatible APIs are a game changer. Switching is mostly a config change, not a rewrite

Quick Code: How I Route Between Models

Heres a simple example of how I handle multiple models through one client:


python
from openai import OpenAI

client = OpenAI(
    api_key="ga_xxxxxxxxxxxx",
    base_url="https://global-apis.com/v1"
)

def get_completion(prompt, difficulty="easy"):
    if difficulty == "easy":
        model = "deepseek-v4-flash"  # $0.25/M
    elif difficulty == "medium":
        model = "
Enter fullscreen mode Exit fullscreen mode

Top comments (0)