semiconductor / news / / Wccftech
DeepSeek V4 is out, bringing major optimizations, including up to 1.6T model sizes.
DeepSeek V4 introduces a 1.6 trillion parameter Pro model and a 284 billion parameter Flash version.
KEY POINTS
- V4 reduces single-token inference FLOPs to 27% and KV cache to 10% for a 1M-token context.
- NVIDIA Blackwell GPUs provide Day-0 support for DeepSeek V4, achieving nearly 3500 TPS per GPU.
- DeepSeek V4 uses FP4 (MXFP4) quantization for faster rollouts and lower memory traffic and latency.
- Huawei's 2026 Ascend 950PR and 950DT chips will support MXFP4, ensuring DeepSeek V4 compatibility.
COMPANIES
Summarized by Newsio from Wccftech. How we summarize →