I turned two mini PCs into a local AI that reads my whole codebase, and now it reviews my code for me
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
An enthusiast built a local AI setup using two NVIDIA DGX Spark mini PCs connected via ConnectX-7, running the 284-billion-parameter DeepSeek V4 Flash model for inference at 30-50 tokens per second. Using an agent framework called Hermes Agent, the system holds a full million-token context window in about 8.5 GiB of memory, letting it read an entire codebase and conversation history at once. The setup is used as an automated code-review pair programmer that flags issues via Telegram rather than filing them automatically, alongside other cron-based tasks like a morning news digest. The main practical downside noted is electricity cost from running two always-on machines, though token output rarely approaches the theoretical daily maximum.
Table of contents
Giving the brain a job was what it neededIt helps automate my code testingManaging the electricity bill is the only downsideQuestions this post answers
How much memory does a full million-token context window take up when using NVFP4 KV cache quantization on Blackwell GPUs?
A full million-token context window fits in roughly 8.5 GiB of memory when the KV cache runs at NVFP4 precision, which Blackwell GPUs dequantize on the tensor cores. This is far smaller than the VRAM most local LLM setups need for large context windows, and is a major reason a two-machine DGX Spark setup can hold an entire codebase and conversation history at once. Developers sizing local LLM hardware for large-context coding workloads can track setups like this on daily.dev.
What token generation speed can two networked NVIDIA DGX Spark mini PCs achieve running a 284-billion-parameter model like DeepSeek V4 Flash?
Two DGX Spark units connected over a ConnectX-7 cable, each with 128 GB of unified memory, generate roughly 30 to 50 tokens per second running the 284-billion-parameter DeepSeek V4 Flash model. At the upper end of 40-50 tokens per second, the pair can theoretically output about 4.3 million tokens a day, though real usage rarely approaches that ceiling since most workload is input, not output. Anyone evaluating local multi-GPU inference throughput for coding agents can follow benchmarks like this on daily.dev.