DeepSeek Adapts V4 to Huawei Ascend Chips

DeepSeek's release notes and Hugging Face model cards confirm what early coverage reported: DeepSeek-V4 launched in preview on April 24, 2026 in two MIT-licensed variants, V4-Pro at 1.6 trillion total parameters with 49 billion activated per token and V4-Flash at 284 billion with 13 billion activated, both with a 1,000,000-token context window. The models use a hybrid attention design combining compressed sparse attention mechanisms to cut compute and memory cost on long-context workloads; DeepSeek states V4-Pro needs roughly 27% of the single-token inference FLOPs and 10% of the KV cache of V3.2 in the 1M-token setting. Reuters first reported the preview release and noted support for agent frameworks and OpenAI/Anthropic-compatible API formats.

TrendForce reports Huawei Ascend, Cambricon, Hygon, and Moore Threads completed day-0 adaptation at launch, with adaptation work completed in April 2026 per Pandaily. SCMP's coverage cites Huawei livestream remarks on early compatibility for Ascend 950PR and Ascend 950DT chips, and eeworld and TrendForce describe full inference-stage Ascend compatibility while citing only rumored, unaudited utilization rates. The pattern matters more than any single number: synchronized adaptation means Chinese deployers no longer wait months for domestic-silicon support after a major open-weight release.

Three concrete implications. First, on-premises options in China now track the open-model frontier with little delay, strengthening the case for domestic-accelerator procurement where NVIDIA supply is constrained. Second, the OpenAI/Anthropic-compatible API surface lowers switching costs for integrators looking to swap or augment hosted proprietary models. Third, million-token context at sharply reduced FLOPs and KV-cache cost, if the claimed efficiency holds in production, moves long-document and agent-trajectory workloads from impractical to routine on mid-sized clusters.

Broader availability of Ascend 950 supernodes in the second half of 2026, which TrendForce and SCMP flag as the trigger for higher throughput and lower Pro pricing; empirical throughput and utilization numbers from early Ascend deployments, since the cited utilization figures are rumored rather than audited; and agent-framework compatibility in practice. No source yet provides audited, end-to-end production benchmarks for V4 on Ascend at scale; vendor statements and livestream remarks dominate the record.