DeepSeek V4 Pro (DeepSeek)

DeepSeek V4 Pro is documented as an open-weight MoE model from DeepSeek under the MIT license, with the current GA checkpoint V4-Pro-0813 (released August 13, 2026) featuring 1.6 trillion total parameters and 49 billion active per token, paired with a 1-million-token context and 384K maximum output. The supplied page e For self-hosting, the excerpt specifies an 8x B200 node running SGLang with the NVFP4 kernel (referenced Lambda pricing $53.52/hr) or a 4x GB300 node with vLLM as reference configurations for V4 Pro, and reports InferenceX serving throughput of 216 tok/s for a single user on B300 with SGLang FP4. The guide, last update

BenchLM's tracker page for DeepSeek V4 Pro 0813 explicitly names the exact checkpoint, confirming its August 13, 2026 release and a 1-million-token context window, with the model ranking 32nd of 232 tracked models at a composite score of 66.4/100 as of September 10, 2026. The excerpt lists category-level rankings: reas The page reports API pricing for V4 Pro 0813 at $0.435 input and $0.87 output per million tokens with cached input at $0.003625 per million (noting a blended $0.65), and measured speed at approximately 69 tok/s with a 30.68-second first-token latency. These figures diverge from the off-peak $0.66/$1.98 tariff cited els