Intel Ships INT4 DeepSeek-V4 Pro and Flash via AutoRound — No MXFP4 Required
Intel has released W4A16 INT4 quantizations of both DeepSeek-V4-Pro and DeepSeek-V4-Flash through its AutoRound tool, with weights available on HuggingFace. Critically, the quantizations are runnable without MXFP4 hardware support — the specialized format that had previously been a bottleneck for wide hardware deployment. This makes high-quality DeepSeek V4 inference accessible on a broader range of server and workstation hardware without requiring next-generation GPU capabilities.
Why It Matters
Combined with the 2-bit GGUF local inference milestone also reported this cycle, Intel's INT4 release reinforces that the DeepSeek V4 family is becoming the de facto benchmark for accessible frontier-quality open-weight inference — with Intel, the broader open-source community, and hardware vendors collectively eroding the datacenter-only requirement.