DeepSeek vs Qwen vs Kimi vs GLM: Which One Should You Actually Use?

Look, deepSeek vs Qwen vs Kimi vs GLM: Which One Should You Actually Use?

I've been routing production traffic through Chinese-hosted LLMs for about eight months now, and I keep getting the same Slack DM from other backend folks: "which one do I actually pick?" Fair question. The marketing pages all claim they're the best. Pricing pages look like they were generated by someone who hates humans. And every benchmark post on Hacker News contradicts the last one.

So here's my honest breakdown after running all four families (DeepSeek, Qwen, Kimi, GLM) through the same Global API unified endpoint for real workloads — customer support summarization, code review bots, document extraction, the usual mid-market SaaS stuff. I'll show you what I measured, what I paid, and where I'd deploy each one tomorrow.

The 30-Second Version (For People Who Skip to the Bottom)

Family	Sweet Spot Model	Output $/M	Vibe
DeepSeek	V4 Flash	$0.25	Cheap, fast, surprisingly good
Qwen	Qwen3-32B	$0.28	The "batteries included" choice
Kimi	K2.5	$3.00	Pays for itself on reasoning tasks
GLM	GLM-5	$1.92	Best Chinese-language output I've seen

If you want the long version with code and gotchas, keep going.

The Full Comparison Matrix

Before I go deep on each one, here's the spec sheet I wish someone had handed me six months ago:

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
Cheapest Option	V4 Flash @ $0.25/M	Qwen3-8B @ $0.01/M	— (all premium)	GLM-4-9B @ $0.01/M
Top Pick	V4 Flash @ $0.25/M	Qwen3-32B @ $0.28/M	K2.5 @ $3.00/M	GLM-5 @ $1.92/M
Code Generation	5/5	4/5	4/5	3/5
Chinese Output	4/5	4/5	5/5	5/5
English Output	5/5	4/5	4/5	4/5
Reasoning	4/5	4/5	5/5	4/5
Throughput	5/5	4/5	3/5	4/5
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 ✅

All of them speak the OpenAI Chat Completions protocol, which means zero migration cost if you're already running OpenAI client libraries. Under the hood, Global API normalizes the auth and base URL so you can swap models by changing one string. That's the part that made this evaluation even possible — I wasn't writing four different SDKs.

DeepSeek: The Default I Keep Recommending

I started with DeepSeek as a cost experiment and ended up routing about 60% of my traffic to it. Here's why.

Model Roster and What I Actually Use

Model	Output $/M	What I Run It For
V4 Flash	$0.25	Daily-driver chat, inline code completions, content rewrites
V3.2	$0.38	When I want the latest architecture but don't need the Pro tier
V4 Pro	$0.78	Production pipelines where output quality matters more than cost
R1 (Reasoner)	$2.50	Math proofs, multi-hop logic, anything with a clear correct answer
Coder	$0.25	Code-specific tasks, refactoring suggestions, test scaffolding

What I Like

• Price-to-quality is genuinely absurd. V4 Flash at $0.25/M output is in the same neighborhood as GPT-4o quality on most of my eval prompts. Not identical — fwiw, GPT-4o is still slightly better at nuanced English — but close enough that the 40x price difference makes it not interesting to argue about.
• It's fast. I'm clocking roughly 60 tokens/second on V4 Flash, which is among the fastest I've measured. If you're building user-facing chat UIs, latency matters more than people admit.
• English output is strong. No obvious "translated from Chinese" artifacts in my testing.
• Code generation holds up. Top-tier on HumanEval and MBPP-style prompts in my own internal suite.

What Annoys Me

• No real vision support. If you need image understanding, DeepSeek isn't the answer (yet).
• Chinese output is a half-step behind GLM and Kimi. Noticeable on literary or culturally-specific prompts.
• Fewer SKUs than Qwen. If you want a very specific size for distillation or on-prem reasons, you're more limited here.

Code: My Default Routing Logic

Here's roughly what my request handler looks like when I want V4 Flash behavior through 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": "system", "content": "You are a concise backend code reviewer."},
        {"role": "user", "content": "Review this diff and flag any race conditions."}
    ],
    temperature=0.2,
    max_tokens=800,
)

print(response.choices[0].message.content)

Notice the base URL — that's the only thing different from a vanilla OpenAI client. The rest is identical. This is what makes model shopping realistic. I can A/B test V4 Flash against GPT-4o by changing exactly one string and a key.

Qwen: When You Need Every Size, Modality, and Price Point

Qwen is the Swiss Army knife. If Alibaba made a model for it, it exists in the Qwen lineup.

Model Roster

Model	Output $/M	What It's Good At
Qwen3-8B	$0.01	Ultra-cheap classification, routing, anything where latency-per-token matters more than depth
Qwen3-32B	$0.28	The general-purpose workhorse — my second-fallback model
Qwen3-Coder-30B	$0.35	Specialized code tasks; better than 32B for code specifically
Qwen3-VL-32B	$0.52	Image understanding, document OCR, screenshot parsing
Qwen3-Omni-30B	$0.52	Audio + video + image in one model (rare to see this)
Qwen3.5-397B	$2.34	The "we need a big brain" enterprise reasoning tier

What I Like

• The range is unmatched. From $0.01/M to $3.20/M, there's a Qwen model for literally any budget constraint I've ever had. This is the family where you can do proper model cascading — start cheap, escalate only when confidence is low.
• Multimodal is real here. Qwen3-VL handles screenshots and documents well; Qwen3-Omni is one of the few models that does audio without making me cry.
• Alibaba infrastructure means uptime. In six months, I've had maybe two real outages on Qwen models through Global API. The SLOs hold.
• Frequent updates. Qwen3.5, Qwen3.6, etc. — they're shipping improvements regularly. imo, this matters more than people give it credit for.

What Annoys Me

• Naming is chaos. Qwen3-Coder-30B vs Qwen3.5-397B vs Qwen3-VL-32B — I have a spreadsheet just to keep them straight. Some standardization would be nice. RFC 9 from me: please pick one naming scheme.
• Mid-range English isn't DeepSeek-tier. It's good, but if pure English quality is the goal, DeepSeek and Kimi edge it out.
• Some models feel overpriced. Qwen3.6-35B at $1/M output is a tough sell when GLM-5 at $1.92/M is right there.

Code: Multimodal Qwen Call

response = client.chat.completions.create(
    model="Qwen/Qwen3-VL-32B",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Extract all line items from this invoice."},
            {"type": "image_url", "image_url": {"url": "https://example.com/invoice.png"}}
        ]
    }],
    max_tokens=1500,
)

I've replaced about 80% of my OCR pipeline with this single call. The $0.52/M is worth it when the alternative is a separate OCR service plus an LLM.

Kimi: The Brain That Costs Real Money

Kimi is the model I pull out when the task is hard enough that I don't want to gamble on a budget model.

Model Roster

Model	Output $/M	Best For
K2.5	$3.00	Hard reasoning, multi-step planning, complex agent loops
(premium tier)	$3.50	The flagship; used sparingly

There isn't really a "cheap Kimi" tier. fwiw, I think that's intentional — Moonshot AI is positioning Kimi as the premium reasoning option, and the pricing reflects it. At $3.00/M output, Kimi K2.5 costs about 12x what DeepSeek V4 Flash costs. So why would anyone use it?

Because sometimes the cheap model gets it wrong, and the wrong answer costs you more than $3.00/M ever will.

Where Kimi Earns Its Premium

• Reasoning benchmarks. K2.5 is the top performer I've tested on multi-step logic, math proofs, and agent planning tasks. When I need a model to think through 5-6 steps of tool use without losing the thread, Kimi is what I reach for.
• Chinese-language depth. Top-tier on Chinese literary, idiomatic, and culturally-nuanced prompts. On par with GLM for this.
• Long-context coherence. Even at 128K tokens, K2.5 doesn't seem to "forget" the middle of the prompt the way cheaper models do.

Where I Don't Use Kimi

• Anything latency-sensitive. It's slower than the other three families. For user-facing chat, the extra thinking time is noticeable.
• Simple classification or extraction. Massively overkill. Use Qwen3-8B at $0.01/M.
• High-volume pipelines. $3.00/M adds up fast. I cap Kimi usage in my routing layer to maybe 5% of total requests.

Routing Pattern That Works

In my service, the logic looks something like:

python