The launch of a pair of new open source, open weights DeepSeek v4 models provides some insight into the technical battle to optimize data center inference hardware and software leverage new model techniques. DeepSeek v4 has added some sophisticated new long context attention optimization approaches, that they have ostensible worked with Huawei to optimize, for a couple of months prior to their first preview.
We don’t get to see inside how the optimization happens with the proprietary frontier model labs, like OpenAI and Anthropic, but the system optimization approaches are quite visible with DeepSeek v4 via both NVIDIA news and updates from the various inference model servers, vLLM, SGLang, etc.
developer.nvidia.com
Also interesting that SemiAnalysis has incorporated DeepSeek-v4 Pro into their data center level inference benchmarking suite within a day or two of availability, showing both unoptimized and optimized results.
inferencex.semianalysis.com
We don’t get to see inside how the optimization happens with the proprietary frontier model labs, like OpenAI and Anthropic, but the system optimization approaches are quite visible with DeepSeek v4 via both NVIDIA news and updates from the various inference model servers, vLLM, SGLang, etc.
Build with DeepSeek V4 Using NVIDIA Blackwell and GPU-Accelerated Endpoints | NVIDIA Technical Blog
DeepSeek just launched its fourth generation of flagship models with DeepSeek-V4-Pro and DeepSeek-V4-Flash, both targeted at enabling highly efficient million-token context inference.
developer.nvidia.com
Also interesting that SemiAnalysis has incorporated DeepSeek-v4 Pro into their data center level inference benchmarking suite within a day or two of availability, showing both unoptimized and optimized results.
AI Inference Benchmarks | InferenceX
Compare AI inference latency, throughput, and time-to-first-token across GPUs and providers. Real benchmarks on NVIDIA GB200, H100, AMD MI355X, and more.