I Got DeepSeek V4 Flash Running at 60 tok/s Locally

Link to Repo - https://github.com/tonyd2wild/DeepSee... Link to Discord -   / discord   Link to Coding Agent - https://github.com/tonyd2wild/coding-... Minimax M3 Repo - https://github.com/tonyd2wild/MiniMax... I finally got DeepSeek V4 Flash running locally on 2 NVIDIA DGX Sparks, and in this video I’m showing the real performance, the repo/recipe, the speed, concurrency, agent workflow, and why this might be the best local AI model setup right now for people running Sparks. We’re talking 50–60 tokens per second, up to 1 million context, a massive KV pool, and real-world agent use — not just benchmarks. I test single stream performance, multiple concurrent sessions, different agent prompts, coding tasks, and how I’m actually using DeepSeek V4 Flash inside my local AI workflow with supervisors, workers, Discord agents, automation, and my larger hybrid cloud/local setup. If you have 1 or 2 DGX Sparks, this is one of the most important models to try right now. DeepSeek V4 Flash is fast, useful, and powerful enough to run real local agents without needing to rely on the cloud for every task. I also compare where it fits next to models like MiniMax M3, GLM, Qwen, and other local model options. In this video: Running DeepSeek V4 Flash on 2 DGX Sparks 1M context local AI setup 50–60 tokens per second performance Single stream and concurrency testing Coding agent demo Local AI automation workflow Discord agent setup DeepSeek vs MiniMax M3 thoughts Why this model matters for local AI builders Repo / recipe will be linked below. If you’re building with local AI, DGX Spark, RTX 3090s, Qwen, DeepSeek, MiniMax, GLM, or agent workflows, make sure you subscribe because we’re testing all of it over here. Join the Discord for local AI, DGX Spark, RTX 3090 builds, recipes, and agent workflow talk. #DeepSeek #DGXSpark #LocalAI #AIModels #DeepSeekV4 #Tech2Wild #NVIDIA #AIAgents #Homelab #OpenSourceAI