DeepSeek V4 Flash (Inference.net)

DeepSeek launched its fourth generation of flagship models with DeepSeek-V4-Pro and DeepSeek-V4-Flash, both built for efficient million-token context inference. DeepSeek-V4-Flash is a smaller mixture-of-experts model with 284B total parameters and 13B active parameters, aimed at high-speed, high-efficiency workloads such as long-context coding, document analysis, retrieval, and agentic workflows. NVIDIA's technical blog details that both models support a 1M-token context window and use hybrid attention, which cuts per-token inference FLOPs by 73% and KV cache memory by 90% versus DeepSeek-V3.2. The post also notes day-0 SGLang and vLLM serving recipes for DeepSeek-V4 on Blackwell and Hopper, with multinode prefill/decode disaggregation scaling to 100+ GPUs and support for tool calling, reasoning, and speculative decoding.