By | World Model Workshop
DeepSeek V4 has once again sent shockwaves across China.
Parameter scale, context length, benchmark scores—these technical metrics have been repeatedly compared in various reports.
But if we only focus on the surface-level data, we miss the most strategically significant core of this release.
Over the past three years, China's large models have been stuck in an awkward reality: training relies on NVIDIA, and inference also relies on NVIDIA. Domestic chips are merely backup options.
If NVIDIA cuts off supply, the entire Chinese model community would be thrown into anxiety.
But today, DeepSeek V4 has demonstrated with strength:
A cutting-edge trillion-parameter large model can also run stably and efficiently on domestic computing power.
The significance of this achievement goes beyond the model's technical metrics itself.
Breakthrough in Domesticization
To truly understand the difficulty of this domestic adaptation, one must first grasp NVIDIA's chip empire.
NVIDIA possesses not just chips, but a highly closed-loop, complete ecosystem:
On the hardware side, there is the GPU chip family, along with NVLink and NVSwitch enabling high-speed networking between chips;
On the software side, CUDA is an AI operating system meticulously built by NVIDIA over more than a decade.
It is like a highly optimized factory, with every layer—from the most basic operators (the fundamental units of model computation) to parallel computing, memory management, and distributed communication—tailor-made for NVIDIA GPUs.
In other words, NVIDIA doesn't just sell engines; it also builds the roads, gas stations, repair shops, and navigation systems.
Almost all global top-tier large models have grown within this ecosystem.
Switching to domestic computing power, however, presents a completely different scenario.
Different hardware architectures, different interconnection methods, varying levels of software stack maturity, and a tool ecosystem still rapidly catching up.
For DeepSeek to adapt to domestic chips, it is not simply a matter of changing engines; it is like switching a race car already speeding on a highway to a mountain road still under construction.
The slightest misstep could lead to jitters, loss of speed, or even a complete halt.
This time, DeepSeek V4 did not choose to merely optimize further along the CUDA path but began simultaneously entering the adaptation chain for domestic computing power's software stack.
Based on public information, V4 has achieved a breakthrough on domestic inference chips, deeply adapting to Huawei's Ascend 950 chip, and Cambricon could also run stably on the day of the model's release, truly achieving Day 0 adaptation.
This means that cutting-edge models are beginning to have the possibility of landing within the domestic chip system.
How did DeepSeek V4 achieve this?
The first step occurred at the model architecture level.
V4 did not choose to force domestic chips to handle the 1M context with brute force but first made the model itself more efficient.
The most critical design in the official technical report is the CSA + HCA hybrid attention mechanism, as well as KV Cache compression and other long-context optimizations.
Simply put, traditional long-context inference requires the model to spread out an entire library to search through every time it answers a question, quickly consuming memory, bandwidth, and computing power.
V4's approach is to first re-index, compress, and filter the materials in the library, sending only the most critical information into the computation chain.
In this way, the 1M context no longer relies entirely on hardware brute force but first reduces the computational and memory costs through algorithms.
This is crucial for domestic chips.
If the model still heavily depends on memory bandwidth and mature CUDA libraries, domestic chips, even if they can run, would struggle to run cheaply and stably.
By reducing the inference burden first, V4 is essentially easing the pressure on domestic computing power.
The second step occurred at the MoE architecture and activated parameter level.
Although V4-Pro has a total of 1.6 trillion parameters, only about 49 billion parameters are activated during each inference; V4-Flash has 284 billion total parameters, with about 13 billion activated each time.
This means it doesn't pull out all parameters for computation every time it is called but functions like a large team of experts, where only the relevant experts are called upon when a task arrives.
For domestic chips, this is equally important.
It reduces the computational pressure that must be borne with each inference, making long-context and Agent scenarios easier to handle on inference cards.
The third step is adaptation at the operator and kernel levels.
The strongest aspect of the CUDA ecosystem is that a large amount of underlying computation has been polished to maturity by NVIDIA, and many high-performance computations can be directly called.
The significance of V4 lies in its extraction of some key computations from NVIDIA's black box, turning them into more transferable and adaptable custom computation paths.
Put more simply, V4 is like disassembling the most critical parts of the engine, allowing manufacturers like Huawei Ascend and Cambricon to readjust them according to their own chip structures.
The fourth step is at the inference framework and service layer.
If domestic chip adaptation only stays at the "running a demo" stage, its industrial significance is limited. What truly deserves attention is whether it can enter a callable, billable service system.
According to internal tests, on Ascend 950PR, V4's inference speed has significantly improved compared to earlier versions, with a notable reduction in energy consumption. Single-card performance in specific low-precision scenarios has reached more than twice that of NVIDIA's specially supplied H20.
DeepSeek officially mentioned that currently, V4-Pro is limited by high-end computing power, with limited service throughput. It is expected that prices will drop significantly in the second half of the year after the batch release of Ascend 950 super nodes.
This indicates that as domestic hardware like Ascend is mass-produced in batches, V4's future throughput and cost-effectiveness will be further optimized.
However, it is worth noting that V4 has not entirely replaced NVIDIA's GPUs and CUDA. Model training may still rely on NVIDIA, but inference can gradually be domesticized.
This is actually a very realistic commercial path.
Training is a阶段性投入, done once, tuned once, iterated once. Inference is an ongoing cost, with millions or billions of user calls daily, each requiring computing power.
The real long-term money burner for model companies will increasingly lean towards inference. Whoever can handle inference demands more cheaply and stably will gain a real advantage in industrial applications.
DeepSeek V4 is the first to present a path for the inference deployment of China's cutting-edge models that does not take NVIDIA CUDA as the default premise.
This step is significant enough.
V4's Impact on Industrial Applications
If domestic chip adaptation answers whether it can run, then price answers another more practical question:
Can enterprises afford to use it?
In the past, DeepSeek's most impressive feat was its ability to deliver near-cutting-edge model capabilities at extremely low prices.
This was true in the V3 and R1 eras, and it remains true with V4.
The difference is that this time, it is not fighting a price war in ordinary context windows but continues to压低价格 under the premise of 1M context + Agent capabilities.
According to DeepSeek's official pricing:
V4-Flash cache-hit input is 0.2 yuan per million tokens, cache-miss input is 1 yuan per million tokens, output is 2 yuan per million tokens;
V4-Pro cache-hit input is 1 yuan per million tokens, cache-miss input is 12 yuan per million tokens, output is 24 yuan per million tokens.
Comparing it with similar domestic models:
Compared to Alibaba's Qwen3.6-Plus in the 256K-1M tier, V4-Pro's output price is about half, and V4-Flash is even lower.
Compared to Xiaomi's MiMo Pro Series in the 256K-1M tier, both V4-Flash and V4-Pro are significantly cheaper.
Kimi K2.6 has a context of 256K; in comparison, V4-Pro has a longer context and lower price; V4-Flash directly pushes high-frequency call costs to another level.
This is hugely significant for enterprise applications.
Because a 1M context means the model can read an entire code repository, thick contract packages, hundreds of pages of prospectuses, long meeting minutes, or the historical state accumulated by an Agent continuously performing tasks in one go.
In the past, many enterprise applications were stuck here: the model's capabilities were sufficient, but the context wasn't; the context was sufficient, but the price was too high; the price was acceptable, but the model's capabilities weren't stable enough.
For example, an enterprise building an investment research Agent needs the model to read annual reports, earnings call transcripts, industry reports, competitor news, and internal minutes simultaneously.
When the context is only 128K or 256K, the system often has to constantly slice, retrieve, and summarize, with information lost in multiple compressions.
A 1M context allows the model to retain more raw materials, reducing omissions and disconnections.
Another example is a code Agent.
It doesn't just write a few lines of code at once but needs to read the repository, understand dependencies, modify files, run tests, and fix issues based on errors. This process repeatedly consumes tokens.
If each step is expensive, the Agent can only be used for demonstrations, but if tokens are cheap enough, it can enter real development workflows.
This is also the industrial value of V4.
It may not be the strongest model, but it could become the most frequently used model in enterprises.
DeepSeek has once again turned AI from a exclusive toy for a few large companies into a productivity tool that can be deployed at scale across industries.
The True Value of V4
When the 1M context reaches the front lines at extremely low prices, the true weight of DeepSeek V4 becomes apparent.
All of this is built on a foundation of domestic computing power that is not yet mature.
Facing the systematic gap in the domestic chip ecosystem, the DeepSeek team did not choose to wait for the ecosystem to mature before launching.
They repeatedly delayed the release window, investing months in deep joint debugging with partners like Huawei. The engineering difficulty of this far exceeded external imagination.
It is precisely for this reason that V4's achievement of near-top-tier closed-source model inference and Agent capabilities on domestic computing power is格外不易.
V4 has proven with itself that even facing the阶段性差距 of hardware ecosystems, Chinese teams can still achieve competitive performance through极致 engineering investment and software-hardware协同 innovation.
Of course, there is still a gap to full maturity.
The完善度 of the Ascend platform's toolchain, the stability of ultra-large-scale clusters, and deep optimization for more vertical scenarios all require continued joint efforts from all parties in the industry.
But the success of V4 has laid a借鉴 path for subsequent models.
It has injected a shot of confidence into the自主可控 of the entire AI supply chain.
In the current external environment full of uncertainties, this resilience to break through amidst limitations deserves more respect than mere parameter metrics.
"Not lured by praise, not frightened by slander, following the path of principle,端正 oneself with integrity."
This quote from DeepSeek's official text is its best footnote.
