七套本地硬件跑 V4.1 Flash 等模型做浏览器世界

FHILY👑 @Oluwaphilemon1原推发布于 2026.10.25

在 X 上打开原推

Seven local AI agent experiments used Qwen3.8 and DeepSeek-V4.1 models to build interactive worlds that people could explore directly in a browser. The hardware ranged from a consumer GPU with 16GB of VRAM to four DGX Spark systems with 512GB of combined memory. The experiments used locally configured open-weight models and inference tools, without cloud API calls for the workflows described. The interesting part is the range of hardware involved. You don’t necessarily need a multi-GPU server to start experimenting with local coding agents. But running larger models, supporting longer contexts and coordinating more demanding workloads can require substantially more memory. The seven configurations 1. RTX 5060 Ti with 16GB VRAM Qwen3.8-27B ran using MiaAI_lab’s inference kit. This represents the consumer-GPU end of the experiments. Running a 27B model on a 16GB card generally requires careful memory management and an appropriate quantization or offloading configuration. 2. RTX 5090 with 32GB VRAM The same Qwen3.8-27B model and inference kit were tested on an RTX 5090. The larger VRAM capacity provides more room for model weights, runtime allocations and context. However, the comparison does not establish a measured speedup over the RTX 5060 Ti because the supplied results do not include matched performance figures. 3. One DGX Spark with 128GB memory Qwen3.8-Flash-Next ran on a single DGX Spark using TensorFold. This moves the experiment beyond a conventional consumer GPU configuration, providing a larger memory pool for local model inference. 4. Two DGX Sparks with 256GB combined memory DeepSeek-V4.1-Flash ran across two DGX Spark systems using TensorFold. Using multiple systems introduces additional engineering considerations, including how model execution is distributed and how communication between devices affects performance. The combined memory capacity alone does not establish the model’s speed or how efficiently the systems are being utilized. 5. Mac with 256GB unified memory running Qwen3.8-Flash-Next This configuration used oMLX to run Qwen3.8-Flash-Next locally on a Mac. Unified memory allows the CPU and GPU to share the system’s memory pool, which can be useful when working with models that would exceed the dedicated VRAM capacity of many consumer GPUs. 6. Mac with 256GB unified memory running DeepSeek-V4.1-Flash The same broad hardware configuration was used with DeepSeek-V4.1-Flash through oMLX. Testing different models on the same class of machine provides another way to explore local deployment, although a direct comparison would require details about quantization, context length and runtime settings. 7. Four DGX Sparks with 512GB combined memory The largest configuration used DeepSeek-V4.1-Flash across four DGX Spark systems, with tensor parallelism. Tensor parallelism distributes parts of a model’s computation across multiple devices. This can make larger workloads possible, but it also introduces communication overhead and synchronization requirements. Having 512GB of combined memory does not mean every task will run proportionally faster than on one system. The outcome depends on the model, workload, implementation and how efficiently computation is distributed. What are these agents actually doing? The objective is not simply to generate a code snippet and stop. The agents are asked to turn a prompt into an interactive environment that runs in a browser. That requires translating an idea into an application, writing the implementation, resolving errors and producing something a person can actually explore. A successful result involves more than plausible-looking code. The application needs to run, respond to user input and behave as intended. This makes browser-playable projects a useful demonstration of coding-agent capabilities because the output can be inspected and interacted with directly. https://x.com/niklaslenz_ai/status/2112666535932117104/video/1