Can DeepSeek V4 Flash Code Like Claude? I Tested It Live on a Mac
I recorded myself giving a local model a full-stack app to build in real time, with none of the failures edited out. The model was DeepSeek V4 Flash, running on a 128GB MacBook Pro, and the run is a fair picture of where local coding actually stands. The video is the whole thing
I recorded myself giving a local model a full-stack app to build in real time, with none of the failures edited out. The model was DeepSeek V4 Flash, running on a 128GB MacBook Pro, and the run is a fair picture of where local coding actually stands. The video is the whole thing uncut. This post is the short version and my take.
The uncut build, in one paragraph
DeepSeek V4 Flash plus the Pi coding agent build a working YouTube title and thumbnail generator, with the twist that the app has to call Codex to make its titles and images rather than any cloud API. In the video you watch the context window overflow early, a bad prompt of mine cost 20 minutes, and a real bug hunt before the thumbnails render. It got there. For contrast I ran the same task through Codex alone and it finished in about 20 minutes. That is the honest gap: a local model can do this, but it needs more steering than a frontier cloud model on the same job. If you want to see the struggle and the fix, the video is where to watch it.
How it fits on a laptop at all
284 billion parameters on a MacBook sounds impossible, so here is the setup that makes it work.
The engine is DS4, or DwarfStar, by antirez, the person who created Redis. It runs exactly one model, DeepSeek V4 Flash, instead of trying to be a general runner. Setup is three steps: clone the repo, download the model files, run make to build. On my 128GB M3 Max I pulled the Q2 two-bit quantized build, which is about 81GB on disk. A bigger machine can take the four-bit version. Two things make a 284 billion parameter model fit at all: the aggressive two-bit quantization, and SSD streaming that caches context on disk instead of demanding everything sit in memory.
DS4 starts a local server on port 8000 with both OpenAI and Anthropic style endpoints, so it plugs into Pi, Claude Code, or Codex. I pointed Pi at it and had it review my WhisperAlone codebase. It read the files, flagged a real problem in the hotkey logic, and suggested fixes. At the time of recording DeepSeek V4 Flash sat at number one on OpenRouter for the week and the month.
Worth knowing: it only runs the DeepSeek V4 Flash files built for it, not any general model, and the SSD streaming is where it slows down when the machine is busy.
Both videos above walk through the build and the setup. If your question is whether local beats the cloud on cost for your workload, the calculator on openclawdc.com will show you.