DeepSeek Adapts V4 to Huawei Ascend Chips

Reporting by Reuters and multiple Chinese industry outlets documents that DeepSeek-V4 has been adapted to run on Huawei's Ascend AI compute ecosystem, with adaptation work completed in April 2026 (Pandaily; Reuters). Reuters reported a preview release of DeepSeek-V4 on April 24, 2026, and noted the model supports a 1,000,000-token context window and two variants - Pro and Flash (Reuters). TrendForce reports that Chinese AI chip vendors, including Huawei Ascend, Cambricon, Hygon, and Moore Threads, completed "day-0" or immediate adaptation at launch; SCMP and eeworld report day-0 or immediate adaptation for Ascend chips (TrendForce; SCMP; eeworld). SCMP's coverage cites Huawei livestream remarks about early compatibility for Ascend 950PR and Ascend 950DT chips.

Reporting by TrendForce and eeworld presents specific model sizing and architecture claims: TrendForce reports DeepSeek-V4-Pro at 1.6 trillion parameters with 49 billion activated per inference, and DeepSeek-V4-Flash at 284 billion parameters with 13 billion activated per inference (TrendForce). Reuters and TrendForce both describe the model family as supporting hybrid and sparse attention mechanisms to enable long-context processing and reduced compute/memory costs for million-token workloads (Reuters; TrendForce). Multiple sources report Ascend compatibility for inference; TrendForce and eeworld state that full inference-stage compatibility with Ascend has been achieved and cite rumored device utilization rates (eeworld; TrendForce).

Editorial analysis: Companies and observers in the Chinese AI stack have emphasized rapid porting of large models to domestic accelerators over the past year. Reporting frames the DeepSeek-V4 adaptation as a milestone for that effort because multiple domestic vendors completed near-simultaneous adaptation, an outcome previously associated mainly with NVIDIA GPUs (TrendForce; Pandaily). For practitioners: closer hardware-software coordination shortens the window between model release and production-ready deployment, which affects choices around latency, cost, and on-premises versus cloud inference strategies.

Observers will monitor three operational indicators reported in the coverage. First, TrendForce and SCMP point to broader availability of Ascend 950 supernodes in the second half of 2026 as a trigger for higher throughput and lower Pro pricing (TrendForce; SCMP). Second, empirical throughput and utilization numbers from early Ascend-based deployments will determine whether the reported compatibility translates to production efficiency; several outlets note rumored utilization metrics but do not offer audited figures (eeworld; TrendForce). Third, compatibility across agent frameworks and API formats-Reuters and TrendForce report DeepSeek-V4 supports agent frameworks and OpenAI/Anthropic-compatible API formats-will shape integrator choices and the ease of replacing or augmenting cloud-hosted proprietary models (Reuters; TrendForce).

none of the scraped sources provides audited, end-to-end production benchmarks on Ascend at scale. Reporting includes vendor statements, livestream remarks, and local-industry measurements; independent third-party performance validation was not presented in the cited coverage (Reuters; Pandaily; TrendForce).