By | World Model Workshop
There is now only one question mark left in China's internet: Will DeepSeek V4 be released or not?
Over the past year, the global model competition has already entered a fierce, head-to-head battle mode.
Giants like OpenAI, Anthropic, and Google have maintained a major iteration cycle of 2-3 months, or even just 1 month, rapidly pushing new versions to the market for validation, iteration, and re-validation.
However, DeepSeek has significantly slowed down its major version updates over the past 15 months. V4 has been delayed multiple times, and its pace has clearly fallen behind, transforming from an industry leader to one that is a step slow.
The anxiety of waiting has evolved into a vortex of suspicion.
In early April, some media suddenly hinted: V4 might be released in the coming weeks, but don't set your expectations too high for DeepSeek V4.
Why would such a long-awaited release come with a preemptive warning?
What exactly happened to DeepSeek this year?
The Burden of Localization
The truth might be hidden in a piece of news that most people overlooked.
In January 2025, a Reuters report revealed a detail: After the release of DeepSeek R1, relevant authorities encouraged DeepSeek to use Huawei's Ascend processors instead of continuing to rely on NVIDIA.
In that context, the word "encouraged" carried far more weight than its literal meaning.
DeepSeek is no ordinary startup; it is the first successful example of China's AI breaking through U.S. technological封锁 (blockade).
This symbolic significance quickly turned DeepSeek from a technology company into a key piece on the chessboard of the national strategy for independent and controllable technology.
Shortly after, in February, Liang Wenfeng attended that highly anticipated symposium for private enterprises.
He was seated in the front row, alongside tech giants like Ma Huateng, Ren Zhengfei, and Lei Jun, standing shoulder-to-shoulder as representatives of the national team for new quality productive forces.
The gears of policy orientation began to turn from there.
According to foreign media reports, DeepSeek did indeed attempt to use Huawei's Ascend 910C chips to train its next-generation model in early 2025.
However, the process encountered multiple technical obstacles: insufficient training stability, frequent crashes in large-scale distributed scenarios, and inter-chip communication speeds falling short of expectations.
Huawei dispatched a team of engineers to DeepSeek's office to provide on-site support, but ultimately failed to resolve the adaptation issues during the training phase.
The result was a compromise: DeepSeek continued to use NVIDIA GPUs for the training phase, while Ascend chips were only used for the inference环节 (segment).
This means that, at least in the core环节 of training, DeepSeek spent nearly a year on trial and error.
But the adaptation work did not stop. In 2026, new developments emerged regarding the V4 version.
According to leaks, DeepSeek did not grant NVIDIA early testing access. Instead, it prioritized giving the pre-release version to Huawei's new-generation Ascend 950PR chips for adaptation.
Simultaneously, to分散风险 (spread risks), it also adapted for Cambricon chips.
However, technical challenges remained significant.
According to media reports, the goal this time was to complete the migration at the underlying code level, moving entirely from NVIDIA's CUDA ecosystem to Huawei's CANN framework, achieving full-chain localization replacement for both training and inference.
According to sources close to the project, the core difficulty of the adaptation work lies in precision alignment—ensuring the model outputs consistent results across different hardware ecosystems, which involves extensive adjustments to the underlying code.
This incurred a time cost.
While global mainstream manufacturers maintained a 2-3 month model iteration rhythm, DeepSeek became increasingly slower. During this period, a significant portion of its technical resources were likely invested in domestic chip adaptation.
After all, there is indeed a generational performance gap, along with differences in ecosystem maturity and toolchain completeness, between domestic chips and NVIDIA. The model adaptation process is incredibly time-consuming.
This marks a clear departure from DeepSeek's initial path, which was purely focused on pursuing model performance improvements.
Linkages at the industry chain level were also happening simultaneously.
In early 2026, the market传出 (circulated) news that Alibaba, ByteDance, and Tencent had placed orders with Huawei for hundreds of thousands of Ascend 950PR chips.
A reasonable speculation is: leading cloud vendors are waiting for the validation results of DeepSeek V4 to assess the practical usability of domestic chips in large-scale AI training.
If V4 succeeds, Huawei's 950PR will transform from a technical sample into a commercially viable product. If it fails to meet expectations, it will essentially map out the current capability boundaries of domestic chips for the industry.
Considering Liang Wenfeng's一贯极高 (consistently very high) standards for model releases—never releasing until expectations are met—the imminent launch of V4 likely means it has passed effectiveness tests on the inference side.
If successfully validated, this would be a critical step for DeepSeek, and indeed for China's entire AI sector, towards independent and controllable technology.
The Cost of Identity Transformation
By proactively undertaking the重任 (heavy responsibility) of validating the domestic computing power ecosystem, DeepSeek's choice makes it resemble more of a national mission-oriented company rather than a purely profit-driven market player.
But the costs of this transformation are obvious: a short-term slowdown in pace, increased pressure on talent retention, and a temporary pause in competitiveness.
According to independent evaluations and community data from March-April 2026, DeepSeek's code generation capability in third-party benchmark tests has been significantly surpassed by the Claude 4 series (Opus 4.6 / Sonnet 4.6).
DeepSeek's multimodal processing capabilities are also primarily limited to text + images, lagging far behind Claude and GPT's performance in image analysis, computer use, and video understanding.
Entering 2026, DeepSeek shifted its product focus towards the more challenging field of Agent system engineering.
Based on current community feedback, DeepSeek is接近 (close to) the first tier in code Agents and Chinese search Agents. However, there remains a clear systems engineering gap compared to top international models like OpenAI and Google in areas such as multi-tool coordination, long-chain task execution, and robustness in real-world environments.
This gap may not necessarily indicate a decline in technical ability, but rather seems like the result of trade-offs made between market competition and national strategy.
The cost at the organizational level is equally apparent.
Starting in the second half of 2025, key members of the DeepSeek core team began to leave.
According to confirmation by LatePost, Wang Bingxuan (core author of the first-generation large model), Guo Daya (core author of R1), Wei Haoran (OCR lead), and Ruan Chong (multimodal lead)相继离职 (left one after another).
Behind these names lies the technical积淀 (accumulation) of DeepSeek from V1 to R1.
The reasons for the departures are complex, but the relative disadvantage of the compensation system is a visible factor.
Headhunters revealed that competitors offered packages that were "two to three times" those of DeepSeek, with some major companies directly offering total compensation in the eight-figure range.
As a startup without external funding (its parent company is幻方量化 (Huanfang Quant)), DeepSeek's salaries, while绝对值不低) (not low in absolute terms), cannot match the equity incentives and valuation premiums offered by market-oriented giants like ByteDance, Alibaba, and Tencent.
Liang Wenfeng has begun promoting company valuation work, clarifying option pricing to give the team more certainty.
But against the backdrop of peers like智谱 (Zhipu) and MiniMax going public and their stock prices soaring, the pressure to retain top talent remains significant.
The DeepSeek of today is陷入 (caught in) a kind of identity模糊 (ambiguity).
It still needs commercialization, it still needs to retain talent, but simultaneously it is burdened with the expectations of domestic adaptation.
The conflict arising from this dual identity is perhaps the deep-seated logic behind DeepSeek's increasing slowness over this past year.
Consequently, market expectations for V4's capabilities are also being adjusted downward.
It might not become the blockbuster, record-breaking model that sweeps across screens once again, but it could be a milestone in industrial significance, proving that China's cutting-edge models can achieve usability within a domestic hardware ecosystem.
The report card for V4 might be more important for the long-term direction of China's AI industry.
