DeepSeek V4 has arrived under an MIT license, but its release comes with a stunning admission: the Chinese lab says it is three to six months behind leading Western AI models. For a company that reshaped the global AI narrative just over a year ago, this is more than a footnote.
The new model, which improves on coding and is designed for Chinese chips, marks a step forward but not the leap many expected. DeepSeek's own technical report acknowledges the gap, and industry sources point to training failures and infrastructure challenges that delayed V4.
What DeepSeek V4 Actually Delivers
DeepSeek V4 brings notable enhancements in code generation and a architecture optimized for Huawei's Ascend chips. The model is available under the permissive MIT license, allowing broad use and modification. However, it remains non-multimodal, a feature postponed due to compute constraints.
The improvements are real but incremental. DeepSeek's engineers describe V4 as a step, not a leap—a candid assessment that contrasts with the hype surrounding its predecessor.
- MIT license permits commercial use and modification, fostering adoption.
- Coding improvements make it competitive for developer tasks.
- Architecture for Chinese chips reduces reliance on NVIDIA hardware.
- Non-multimodal limits its versatility compared to Western models.
- Self-reported lag of 3-6 months behind frontier models.
The Delays and Technical Hurdles
V4's launch was delayed by a serious training failure in mid-2025 during a migration from NVIDIA to Huawei Ascend chips. According to sources cited by 36Kr, internal disagreements over technical direction and stringent conditions imposed by founder Liang Wenfeng further slowed progress.
The result: months of setbacks and a model that lacks multimodal capabilities, which are increasingly standard in Western AI systems. The compute shortage has forced DeepSeek to prioritize core language tasks over advanced features.
Why DeepSeek's Symbolism Matters
In January 2025, DeepSeek became a symbol of China's AI ambitions. Its breakthrough model shook markets, challenged the logic of the US tech stock rally, and convinced many that China could compete at the frontier at a fraction of the cost.
V4 complicates that narrative. It shows that maintaining a leadership position is harder than making a splash, especially when geopolitical and hardware constraints come into play.
How V4 Compares to Western Models
DeepSeek's own assessment places V4 behind top Western models like GPT-5 and Gemini Ultra. The table below summarizes key differences based on available information.
| Feature | DeepSeek V4 | Western Frontier Models |
|---|---|---|
| License | MIT (open) | Proprietary (mostly) |
| Multimodal | No | Yes |
| Hardware Optimization | Huawei Ascend | NVIDIA GPUs |
| Performance Gap | 3-6 months behind | Leading edge |
What This Means for the AI Race
DeepSeek's admission highlights the challenges of competing at the AI frontier without access to the most advanced chips. The US export controls on advanced semiconductors have forced Chinese labs to innovate on less powerful hardware, which impacts training efficiency and model capabilities.
Yet DeepSeek's open-source approach could still democratize AI development globally. By releasing V4 under MIT, it enables researchers and startups to build on its work, potentially accelerating innovation outside the US-China binary. The background to this story is set out in China's AI Competition Threatens Anthropic's $2 Trillion IPO Valuation. A related thread runs through Chinese AI Models Qwen3.8-Flash and DeepSeek V4-Flash Disrupt Pricing. Earlier coverage in DeepSeek V4 Ignites China's Domestic AI Chip Industry shows how the argument developed.
The Road Ahead
DeepSeek plans to address multimodal capabilities once compute constraints ease. The lab's ability to secure advanced chips or optimize for domestic alternatives will determine its future competitiveness. A closer look at the same subject appears in Jensen Huang Says AGI Has Arrived With GPT-6 Astra. Anyone following the wider picture can start with One Nation Leader Warren Pickering Admits Drug Use, Drink Driving. The same theme surfaces in DeepSeek V4-Flash: The Cheapest AI Model on the Market Right Now, which takes a different angle.
For now, V4 serves as a reality check: even the most promising challengers face significant hurdles in the fast-moving AI race. Context on how this played out before sits in Trump Press Secretary Bar Exam Controversy: China Admission Raises. A companion piece, DeepSeek Seeks $45B Valuation With First Outside Funding Round, approaches the same subject from another angle.
FAQ
What is DeepSeek V4?
DeepSeek V4 is the latest AI model from Chinese lab DeepSeek, released under the MIT license. It features improved coding capabilities and is optimized for Huawei Ascend chips, but lacks multimodal functionality.
Why is DeepSeek V4 considered behind Western models?
DeepSeek's own technical report states that V4 is three to six months behind leading Western AI models. This gap is attributed to compute constraints and delays in training infrastructure migration.
What challenges did DeepSeek face in developing V4?
DeepSeek encountered a serious training failure during a migration to Huawei Ascend chips, leading to months of delays. Internal disagreements and founder-imposed conditions also slowed progress.
Is DeepSeek V4 open source?
Yes, DeepSeek V4 is released under the MIT license, which allows for free use, modification, and distribution, including for commercial purposes.