DeepSeek-V4 is open source with a 1M context but limited compute. Through algorithmic restructuring and sparsity design, the Pro version approaches top closed-source models; the Flash version activates only 13B parameters, adapting to mid-range ecosystems and lowering the barrier to use.
On April 24th, a shoe finally dropped in China's large model race. DeepSeek-V4 Preview officially launched and was open-sourced simultaneously, making 1M (one million token) ultra-long context a standard feature out of the box for its official service.
If this had happened a year ago, long-text processing capability of this caliber was still a proprietary privilege locked behind enterprise paywalls by leading overseas tech giants. Now, it has been laid directly on the table of the open-source community, becoming infrastructure that developers can use at will. For developers who stay up late wrestling with massive codebases or complex legal contracts, this is undoubtedly good news.
But beneath this democratization of technology, the official announcement retained a remarkably restrained disclosure: "Due to constraints on high-end compute, DeepSeek-V4-Pro's current service throughput is quite limited."
For those accustomed to vendors boasting about their compute reserves at launch events, this bluntness carries a rare sense of sobriety. In the second half of the large model race, everyone in the industry knows exactly how much high-end hardware leverage each player holds. Rather than sustaining the illusion of parameter-level prosperity, it's better to lay the industry's realities bare. DeepSeek's move this time represents a surrender of the obsession with pure benchmark competition, finding a compromise solution that balances core algorithmic breakthroughs, the domestic heterogeneous compute ecosystem still under development, and the real commercial environment of enterprises, while accommodating both technological evolution and hardware constraints.
China's AI industry is shedding its early-stage blind money-burning mentality and entering an extremely pragmatic era of "compute ledger" accountability.
How Does the Pro Version Balance Its Compute Ledger?
Let's look specifically at the V4-Pro, whose throughput has been explicitly throttled. As the flagship of the family, V4-Pro boasts a massive 1.6T total parameter count, but only needs to activate 49B parameters during inference. This extreme sparsity design is not merely a showpiece model for display; under the rigorous scrutiny of real production pipelines, its technical foundation demonstrates formidable resilience.
The ability to handle complex code and logical reasoning is the litmus test for whether a large model can truly enter core production environments. In Agentic Coding evaluation environments, V4-Pro's practical performance firmly places it in the top tier of current open-source models. DeepSeek has already integrated it into its internal code pipeline, making it a heavily relied-upon productivity tool for frontline engineers. Feedback from R&D personnel indicates that its code generation and error correction experience is superior to Sonnet 4.5, approaching Opus 4.6 in non-deep-thinking scenarios, though it still lags behind Opus 4.6's thinking mode.
Behind this practical performance lies the research team's extreme pursuit of algorithmic depth. In world knowledge benchmarks that test pre-training data cleaning quality and knowledge density, V4-Pro leads most existing open-source models, currently trailing only the top closed-source model Gemini-Pro-3.1. As for mathematics, STEM (science, technology, engineering, mathematics), and competitive coding benchmarks, it has earned the right to compete alongside world-class closed-source tech giants.
Achieving this capability is clearly not a matter of simply stacking compute cards. Domestic teams are well aware that competing on high-end GPU reserves is unrealistic. V4-Pro's ability to handle 1M ultra-large contexts under limited VRAM is underpinned by the R&D team's deep reconstruction of the attention mechanism. They implemented a novel attention compression scheme that performs aggressive compression at the token dimension, paired with their signature DSA sparse attention technology (DeepSeek Sparse Attention).
This original technical approach, combined with the first-time introduction of KV Cache sliding windows and compression algorithms, effectively controls the computational overhead and memory footprint of long-sequence processing. To enable developers to actually leverage these capabilities in production, the R&D team made targeted bottom-layer adaptations for mainstream Agent tools such as Claude Code and OpenClaw. The technical documentation even explicitly states that developers can directly enable thinking mode when handling complex tasks by setting the reasoning_effort parameter to max. This kind of systematic engineering optimization under constrained computational resources precisely demonstrates to the industry that even with limited high-end compute, domestic teams can still push the boundaries of model performance through native architectural design.
Who Does the 13B Activation Cap Affect
Those fixated on the Pro version's throughput bottleneck often overlook the commercial pillar hidden behind it by DeepSeek—the Flash version. Some voices in the industry dismiss it as a compromise born of compute shortages, but such views clearly underestimate the management team's long-term considerations. This is a pragmatic positioning move targeting the downstream ecosystem, executed after rigorous cost accounting.
According to publicly available adaptation code information, the Flash version maintains a massive total parameter count of 284B, but its activation parameter count has been precisely capped at 13B.
13B, in a context where peers are trying to push parameters toward the trillion scale, hardly seems noteworthy. But this precisely illustrates the economic logic of Mixture of Experts (MoE) architecture in commercial deployment: total parameters determine the breadth of model knowledge, while activation parameters directly determine the electricity costs and memory bandwidth the server must expend for every API call.
Compressing activation to 13B effectively strips the large model away from expensive top-tier AI computing centers. Its demands on single-card VRAM and peak compute are remarkably restrained. Test results show that the Flash version maintains stable response speed and accuracy when handling massive, high-frequency, simple daily tasks, with no significant degradation in underlying general reasoning capability. For small and medium-sized developers and long-tail enterprises processing tens of thousands of API calls daily, this is a genuinely affordable, runnable productivity tool.
The deeper industry logic lies in the fact that China's mainstream heterogeneous compute chips are still in a catch-up phase in terms of absolute single-card performance. Systems carrying fully activated models are prone to hitting memory walls, resulting in poor operational efficiency; but with the Flash version's mere 13B activation, these chips can run smoothly at low to medium power consumption.
With this move, DeepSeek has revitalized a vast amount of idle mid-to-low-end compute resources in China, providing a highly compatible testing ground for domestic chips desperately in need of deployment scenarios. This infrastructure-building logic of downward compatibility is far more aligned with current commercial realities than merely climbing benchmark leaderboards.
Can Domestic Chips Handle It?
What sparked extensive industry discussion about this release is its full-stack domestic deployment label. For a long time, there has been a disconnect between algorithm companies and domestic chip manufacturers: model vendors worry that immature hardware ecosystems will slow down R&D progress, while chip vendors lack access to the most cutting-edge large models for deep optimization. This time, the deadlock has been substantively broken.
Huawei Computing quickly issued a statement confirming that its Ascend supernode full product line fully supports the new model. From a technical detail perspective, Ascend's underlying chips rely on fused kernels and multi-stream parallelism to effectively reduce system computational overhead, thereby stabilizing inference performance in long-text scenarios. Cambricon also swiftly completed Day 0 adaptation and open-sourced its underlying code, while Hygon's DCU simultaneously announced a closed loop.
But we need to look past the surface of ecosystem prosperity and examine the real friction in the data center when hardware and software are stitched together. Taking the Ascend 950 series chip as an example, according to industry sources, the chip features 112GB of proprietary HBM, 1.4TB/second bandwidth, and a single-card power draw of 600 watts. At specific inference precisions (such as FP4), its single-card compute has already demonstrated extremely strong data performance, reaching 2.87 times that of NVIDIA's H20. However, in the higher-requirement FP16 or FP32 general-purpose training precision ranges, the performance gap between domestic hardware and NVIDIA remains.
Furthermore, the so-called "Day 0 adaptation" still requires crossing hidden costs arising from supply chain opacity before achieving lossless enterprise-grade operations. Supernode hardware's high-speed interconnect standards are extremely closed, and the flow of core components resembles a black box of information. This procurement-side barrier undoubtedly makes large-scale deployment and maintenance of compute systems more complex.
Additionally, this system currently relies heavily on centralized procurement orders from a very small number of large domestic institutions. The scarcity of overseas orders means this compute breakthrough can only circulate within the domestic loop. This monolithic commercial loop means the operational efficiency of the entire software-hardware co-developed system urgently needs refinement through a more diversified commercial environment.
The tight ramp-up of high-end compute capacity has directly led DeepSeek to admit in its announcement that significant price reductions for the Pro version will have to wait until supernodes hit the market in bulk in the second half of the year. Large models and domestic chips have indeed achieved preliminary physical interlocking, but under technical gaps and supply chain constraints, this forward charge while wounded is precisely the most authentic survival snapshot of the domestic compute ecosystem.
Can the Technology Keep Turning if People Leave
Stepping back into the real commercial competition, the arrival of DeepSeek-V4 is an extremely precise strategic defense. Over the past six months or more, this company has been under intense pressure. The C-end arena has turned into a red ocean, with leading players deploying massive capital on intensive advertising. QuestMobile data presents a clear competitive landscape: as of March 2026, Doubao had 345 million monthly active users, Qwen had 166 million, and DeepSeek held its ground with 127 million.
External traffic competition is fierce, and internally, the technical bench is also facing attrition. Poaching competition in the industry has intensified, with key personnel from multiple business lines leaving in succession. According to public resumes and industry information, the core author of the first-generation large language model has confirmed a move to Tencent, a core contributor to V3 has gone to Xiaomi, a core researcher on R1 has joined ByteDance, and core strength in the multimodal direction has also confirmed new destinations. According to industry rumors, Wei Haoran, the core author of the OCR direction, has also resigned.
Changes among core R&D members inevitably invite strict external scrutiny of the company's R&D momentum: will the innovation capability of its underlying architecture be affected?
At this juncture, the release of the V4 preview serves as the most direct response. It proves to the market that the company has established a systematic R&D pipeline with risk-resistance capabilities. Even amid personnel restructuring, its technological evolution logic continues to operate with precision. This organizational resilience built on engineering systems quickly drew positive feedback from the capital markets.
Recently, DeepSeek was reported to be seeking financing at a valuation of no less than $10 billion, planning to raise funds to shore up reserves. According to industry media citing sources close to the transaction, market rumors suggest that a major internet giant is expected to inject capital, potentially pushing this round's valuation even higher. If this deal is finalized, it would rewrite the valuation record in China's large model track, surpassing the previous performance of Moonshot AI. During this critical phase of financing negotiations, delivering substantive results like million-token context and full-stack domestic adaptation is the management's rational move to stabilize the strategic landscape and address external doubts.
Final Thoughts
In a tech business landscape where concepts change rapidly, teams willing to focus on underlying infrastructure building have always been scarce. The release of DeepSeek-V4 has established a pragmatic and sober tone for the second half of the large model competition.
Facing compute bottlenecks, they chose not to gloss over them but instead laid the real supply-demand dynamics of domestic high-end hardware bare for the market; facing downstream deployment needs, they used the 13B-activation Flash version to create survival space for domestic compute chips still in their catch-up phase; facing external traffic encirclement and talent competition, they responded at the industry level with concrete long-text processing capabilities.
The quote from Xunzi cited by the official account on launch day carries profound meaning: "Not swayed by praise, not fearful of slander, follow the path and stand upright."
Models can be open-sourced, but compute will never be free. What DeepSeek delivered this time is not a more powerful model, but a solution for how capabilities can be redistributed when compute becomes a constraint. In a reality where compute remains imperfect, this may be an evolutionary direction closer to the essence of the industry.