What is DeepSeek V4, released as a preview on April 24, 2026, and why does it matter despite being unlikely to replicate the global shock caused by DeepSeek R1—particularly regarding its open-source availability, V4-Pro and faster, cheaper V4-Flash versions, reasoning capabilitie

DeepSeek V4 is DeepSeek’s open-weight, preview-generation model family: a large frontier-oriented model (V4-Pro) and a smaller, faster model (V4-Flash). It is unlikely to recreate R1’s surprise because the market now expects capable, low-cost Chinese models; its larger importance is that it combines competitive model capability with a practical route toward a Chinese-controlled AI software-and-hardware stack. 1415

• What was released: On April 24, DeepSeek released V4-Pro (1.6 trillion total parameters, 49 billion active per token) and V4-Flash (284 billion total, 13 billion active), both mixture-of-experts models with open weights and a 1-million-token context window. 1410

• Two distinct products: Pro is the flagship for harder reasoning, coding, agentic work, and claimed near-frontier performance; Flash activates far fewer parameters, targeting lower latency and cost for high-volume deployments. DeepSeek makes both available as downloadable weights and through its API, rather than limiting access to a closed hosted service. 1415

• Performance and price claims need qualification: DeepSeek says Pro rivals leading closed models and emphasizes reasoning, knowledge, coding, and agent benchmarks. Those are vendor claims in a preview release—not yet a substitute for broad independent testing across real workloads. 1415

• Why the long context is meaningful: V4’s hybrid/selective-attention design treats recent context in detail while compressing older information into a smaller memory representation. DeepSeek and Hugging Face report that, at a 1-million-token context, Pro needs 27% of V3.2’s per-token inference FLOPs and 10% of its KV-cache memory; Flash is reported at 10% of the FLOPs and 7% of the cache. That could make persistent coding repositories, long research corpora, and long-running agents materially more economical. 16

• Developer/agent angle: DeepSeek positioned V4 for coding and tool-using agents, including integration with developer-agent environments such as Claude Code, OpenClaw, and CodeBuddy. In practice, compatibility does not establish parity with Claude or other leading closed models; it means developers can try a much cheaper, self-hostable or API-served alternative in those workflows. 14

• Why it matters geopolitically: V4 is DeepSeek’s first major model adapted for Huawei Ascend hardware, rather than being principally tied to Nvidia GPUs. This is strategically important under US chip-export restrictions because it helps validate a domestic model–chip–software path and supports China’s aim for a more self-reliant AI ecosystem. 23

• But Nvidia is not displaced: A model running on Ascend is not equivalent to replacing Nvidia. Nvidia’s advantage includes mature GPU hardware, interconnects, the CUDA programming ecosystem, optimized libraries, developer familiarity, and global cloud deployment. The available public reporting does not establish what share—if any—of V4’s full training run used domestic chips, as distinct from optimization and inference support; claims of wholly domestic training should therefore be treated as unverified. 23

• Why Ascend 950 could matter: Reuters reported strong demand for Huawei Ascend 950 chips after V4’s launch. If Huawei’s supernode-scale systems provide usable performance, software tooling, supply, and lower total cost, they could reinforce V4-style efficiency gains and create a viable parallel Chinese AI infrastructure. That is a plausible strategic trajectory, not a proven replacement for the Nvidia-based global stack. 1

In short, R1 changed perceptions by showing that a Chinese lab could produce strong reasoning cheaply. V4’s potentially deeper effect is operational: open models, very long context at lower inference cost, agent/coding deployment, and a credible attempt to make those capabilities run on domestic Chinese infrastructure.