Here's a telling number: at the start of 2025, DeepSeek and Qwen together held less than 1% of the global AI market share. By January 2026, that figure had jumped to 15%.
In one year, they went from fringe players to a position the industry can no longer ignore — a remarkably fast ascent.
What happened behind the scenes is worth examining.
DeepSeek V4: The Trillion-Parameter Giant
The full version of DeepSeek V4 has not yet been officially released (as of early April), but various reports and technical details have already painted a clear picture.
Core specs: approximately one trillion total parameters, with about 32 billion activated per token — a classic Mixture-of-Experts (MoE) architecture that delivers far higher computational efficiency than pure dense models.
Several standout features:
Context window expanded to 1 million tokens. V3 had 128K; this is an 8x leap, allowing a single session to hold an entire codebase or a thick book. For tasks like cross-file code analysis or processing ultra-long contract documents, this change is substantial.
Native multimodality. Text, images, video, and audio are processed uniformly without additional adaptation layers. This architectural approach is much cleaner than later-patched multimodal solutions.
Pricing: $0.14 per million input tokens. Roughly one-twentieth the cost of GPT-5.
One case study puts it plainly: a company processing 50,000 financial documents daily spent $12,000 per month on GPT-5. Switching to DeepSeek V4 with comparable accuracy cost only $600.
Qwen 3.5/3.6: Alibaba's Other Card
While DeepSeek V4 was sharpening its edge, Alibaba released several new Qwen versions in quick succession from late March to early April.
Qwen3.5-Omni launched on April 1, and Qwen3.6-Plus was officially announced on April 2, featuring approximately 39.7 billion active parameters, a million-token context window, and built-in reasoning chains.
Key differences from DeepSeek V4:
| DeepSeek V4 | Qwen 3.6-Plus | |
| License | DeepSeek License (restricted) | Apache 2.0 |
| Multilingual support | ~50 languages | 201 languages |
| Parameter scale | ~1T total / 32B active | 397B total / 17B active |
| Enterprise deployment cost | Low | Lower |
Apache 2.0 is a tangible benefit for enterprises — no need to have a legal team review the license, and deployment can proceed directly.
Additionally, support for 201 languages is a major draw for companies with overseas operations or multi-region businesses.
The Open-Source Value Proposition Has Arrived
Honestly, a year ago if you told a business decision-maker to replace GPT-4 with an open-source model, you'd likely be seen as taking a risk. That conversation has now changed.
Several real shifts are underway:
The capability gap is narrowing. On standard tasks like code generation, document processing, and data analysis, DeepSeek V4 and Qwen 3.5 now perform on par with top closed-source models, and even surpass them on certain specialized benchmarks.
The price gap is widening. DeepSeek V4's API price is about 5% of GPT-5's, and Qwen series pricing via platforms like SiliconFlow is in the same ballpark.
Deployability has improved. Both models support local deployment, finally offering a reliable option for compliance requirements that mandate data not leaving the country.
Of course, open source isn't a panacea — for highly specialized tasks (e.g., medical, legal) or scenarios requiring continuous updates and maintenance, closed-source models still hold an advantage. But for the vast majority of everyday AI applications, the math is now worth a serious second look.
What to Watch Next
When the full version of DeepSeek V4 will be released is one of the most anticipated launches this year.
If the full version delivers on its expected performance, combined with its aggressive pricing, the pressure on GPT-5 and Claude Opus series will be direct and intense.
Anthropic recently faced developer complaints over the OpenClaw issue. If the open-source camp raises its competitiveness another notch at this moment, the timing is quite interesting.
Sources: CocoLoop, DeepSeek V4 and Qwen 3.5: Open-Source AI Is Rewriting the Rules in 2026 (Particula Tech), AI Updates Today (April 2026) – Latest AI Model Releases (llm-stats.com)