DeepSeek V4.1 Flash 在 7 种本地硬件配置上跑浏览器世界生成
从 16GB 显存的 RTX 5060 Ti 到 4 台 DGX Spark（512GB、TP4），用 TensorFold、oMLX 等推理方案让本地 agent 生成可在浏览器游玩的世界，DeepSeek 参与其中 3 组配置。
Qwen3.8-27B and DeepSeek-V4.1-Flash just powered a series of local AI agent experiments across seven hardware configurations, ranging from a 16GB GPU to systems with 512GB of memory. The task? Let local AI agents build entire worlds that you can explore directly in your browser. Here are the seven setups: 1. RTX 5060 Ti, 16GB VRAM Qwen3.8-27B using MiaAI_lab’s inference kit. 2. RTX 5090, 32GB VRAM Qwen3.8-27B using the same kit. 3. One DGX Spark, 128GB memory Qwen3.8-Flash-Next running on TensorFold with MiaAI_lab’s recipe. 4. Two DGX Sparks, 256GB memory DeepSeek-V4.1-Flash running on TensorFold. 5. Mac, 256GB unified memory Qwen3.8-Flash-Next running through oMLX. 6. Mac, 256GB unified memory DeepSeek-V4.1-Flash running through oMLX. 7. Four DGX Sparks, 512GB memory DeepSeek-V4.1-Flash using Tensor Parallelism 4 (TP4). The interesting part is the range of hardware involved. A consumer GPU with 16GB of VRAM sits at one end, while multi-system configurations with hundreds of gigabytes of memory sit at the other. Different model architectures, memory capacities, inference engines, and parallelization strategies make these experiments possible. It also demonstrates how much the local AI ecosystem has expanded. You can now experiment with capable open-weight models without relying entirely on cloud APIs. Models by Alibaba’s Qwen team and DeepSeek, powered by community-built inference recipes. Every world is playable in your browser. Open the individual pages to see what each configuration produced. https://x.com/niklaslenz_ai/status/2112666535932117104/video/1