If you have been following the AI industry recently, you may have noticed something interesting: two Chinese open-source models are competing for the same market, and both are challenging closed-source giants that charge tens of dollars per million tokens — at a fraction of the cost.
DeepSeek V4 and Qwen 3.5 (already labeled Qwen 3.6-Plus in some channels) — have you sorted them out?
First, the Numbers
DeepSeek V4:
- Total parameters: approximately 1 trillion (approximately 32 billion activated)
- Context window: 1 million+ tokens
- API pricing: approximately $0.14 per million input tokens
- Training cost: single-digit millions of USD
Qwen 3.5/3.6-Plus:
- Total parameters: approximately 397 billion (approximately 17 billion activated)
- Context window: 1 million tokens
- License: Apache 2.0 (fully open for commercial use)
- Languages supported: 201
For comparison, GPT-5's API pricing is about $5 per million input tokens. DeepSeek V4 costs roughly one-thirty-fifth of that.
Which One Is More Capable?
Honestly, each has its own strengths — they are not the same thing.
Where DeepSeek V4 excels:
- Massive context handling: throw a multi-million-character codebase or long document at it, and it won't break
- Native multimodal: text, images, video, audio all in one — no external adapters needed
- Currently the strongest open-source model for understanding large-scale code projects
Qwen 3.5's advantages:
- Apache 2.0 license: enterprises can freely use it commercially — something DeepSeek cannot offer
- 201-language support: if you are building multilingual applications, especially for Southeast Asian or niche language markets, Qwen is the first choice
- Built-in thinking/non-thinking inference modes: more flexible for agent development
- MathVision benchmark score of 88.6: strong mathematical reasoning
In a nutshell: DeepSeek V4 is for large-document and large-codebase scenarios; Qwen 3.5 is for multilingual, agent development, and compliance-heavy use cases.
Why This Matters
Two years ago, Chinese AI models held only a tiny fraction of the global market. Now, DeepSeek and Qwen together have captured roughly 15% of the global AI market share, and the curve is still climbing.
More importantly, they have broken the default assumption that you need a big budget to get a good model.
GPT-5 is estimated to have cost over $100 million to train, while DeepSeek V4's training cost is likely in the single-digit millions. In the end, the performance gap on many tasks has narrowed to within statistical noise.
What This Means for Developers
If you are choosing an AI API to build a product:
- Budget-sensitive, high call volume: DeepSeek V4's price advantage is too obvious to ignore
- Need a commercially friendly license: Qwen's Apache 2.0 is the cleanest among open-source models
- Building multilingual SaaS: Qwen's 201-language support is genuinely useful
- Large codebase analysis: DeepSeek V4's million-token context plus native code capability is a better fit
Closed-source models still have an edge in complex multi-step reasoning and top-tier programming tasks, but the gap is closing fast.
Not Over Yet
The full version of DeepSeek V4 (trillion parameters) has not actually been officially released yet. What came out on March 9 was the V4 Lite version with approximately 200 billion parameters. The full version is expected within the next few weeks.
On the Qwen side, 3.6-Plus went live on OpenRouter in late March and was officially announced on April 2. It has already posted results surpassing Claude Opus 4.5 on real-world terminal tasks.
Both are pushing hard simultaneously, competing for the same market, and driving prices to the floor — that is the real reason the entire industry is on edge.
Sources: DeepSeek V4 and Qwen 3.5: Open-Source AI Is Rewriting the Rules in 2026 (Particula.tech); CocoLoop, DeepSeek V4 Release Imminent — China's Next-Gen AI Model Race Accelerates (Jangwook.net)