DeepSeek V4: A Step Forward, But Not a Leap
DeepSeek has published its V4 model under the MIT license, featuring notable improvements in code and architecture specifically designed for Chinese chips. However, the developers have admitted in their technical report that they are currently lagging three to six months behind leading Western models. For a lab that transformed the global AI narrative merely a year ago, this delay marks a significant setback.
The Importance of DeepSeek
In January 2025, DeepSeek emerged as a symbol of China’s capabilities in AI, shaking global markets and challenging the American technology narrative by demonstrating that China could compete at the forefront of AI development at a fraction of the cost. Yet, the release of V4 complicates this narrative, as internal engineers describe the model as a step rather than a revolutionary leap forward.
Context and Development Delays
The launch of V4 has faced unexpected delays. Reports suggest that DeepSeek encountered a crisis during a critical training phase in mid-2025 when transitioning from NVIDIA to Huawei’s Ascend chips. Diverging internal views on technical direction led to difficulties in executing founder Liang Wenfeng’s vision, resulting in a model that is not yet multimodal—a feature postponed due to constraints on computing capacity and financial resources.
Innovative Architectural Features
One of the most compelling aspects of V4 lies in its architecture. The model introduces TileLang, a domain-specific language that allows for low-level code to be decoupled from CUDA, NVIDIA’s industry-standard platform. This decoupling aims to enable compilation for different hardware architectures. Furthermore, V4 incorporates MegaMoE, a kernel designed to reduce latency in expert parallelism, built to operate effectively on Ascend hardware. Despite these advancements, the model’s training continues to utilize NVIDIA GPUs, highlighting that true independence remains an aspiring goal rather than a fully realized achievement.
Shifting Market Dynamics
While DeepSeek has been focused on refining its technology, the Chinese AI market has rapidly evolved without it. Notable developments include ByteDance’s Doubao, which has become China’s most downloaded chatbot, and the successful IPOs of MiniMax and Z.ai. Additionally, Alibaba has gained traction through its vertical applications. DeepSeek’s reluctance to produce consumer-focused products has left it at a disadvantage as the market has advanced.
Talent Drain and Funding Issues
Internal challenges have also impacted DeepSeek’s efforts, including the loss of key personnel to competitors like Tencent, ByteDance, and Xiaomi. Liang Wenfeng’s refusal to cede 20% equity to a large investor has further complicated matters. For the first time, DeepSeek is now looking to open an external funding round, a move indicative of its pressing need for resources.
A Blow to the Open Source Narrative
The open-source narrative surrounding Chinese AI as a viable alternative to Western closed models has taken a significant hit. An employee from Qwen publicly stated that “the golden age of nonprofit AI development is over,” indicating a shift in sentiment within the industry.
The Big Question: Regaining Ground
The crucial question now is whether DeepSeek can recover lost ground in the fiercely competitive AI landscape. A lot hinges on Huawei’s Ascend 950, which promises compatibility with V4. However, its production capabilities, with 750,000 units equating to roughly a week of American output when adjusted for quality, pose a monumental challenge. This gap will not be closed solely with innovative architectures but requires tangible advancements in silicon production.