DeepSeek V4 launches optimized for Huawei, not Nvidia
SkimNews Take
DeepSeek's V4 model, by significantly extending prompt length capacity, could enable more complex, multi-turn AI interactions that were previously impractical due to context window limitations.
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- DeepSeek released a V4 preview on April 24 — its first major model since R1 in January 2025 — open-source in two versions: V4-Pro for coding and agent tasks, and V4-Flash for cheaper, faster inference
- V4-Pro is priced at $1.74 per million input tokens and $3.48 per million output tokens, while V4-Flash costs about $0.14 and $0.28 respectively, with both versions offering a 1-million-token context window and reasoning modes
- DeepSeek says V4-Pro matches Anthropic's Claude-Opus-4.6, OpenAI's GPT-5.4, and Google's Gemini-3.1 on major benchmarks, and beats open-source rivals Alibaba's Qwen-3.5 and Z.ai's GLM-5.1 on coding, math, and STEM
- V4-Pro uses only 27% of the computing power and 10% of the memory that V3.2 required for a 1-million-token context, thanks to a redesigned attention mechanism that compresses older text and prioritizes the passages most likely to matter
- DeepSeek gave prerelease access to V4 only to Chinese chipmakers like Huawei — not to Nvidia or AMD — and Huawei confirmed its Ascend 950 supernodes will support the model when they ship at scale in the second half of 2026
- Huawei's Ascend chips still trail Nvidia for training but work for inference, per anonymous sources and Tsinghua professor Liu Zhiyuan cited by MIT Technology Review, who noted DeepSeek appears to have adapted only part of V4's training to domestic silicon
Why it matters: For developers, V4-Pro delivers frontier-level coding at $1.74 per million input tokens — a fraction of OpenAI and Anthropic pricing — while staying open-source. For China's AI stack, V4 is the first top-tier model running on Huawei Ascend chips, with further price drops explicitly tied to Ascend 950 supernode shipments in H2 2026.
