DeepSeek Harness setup: local vs hosted models

DeepSeek Harness is an open source coding assistant built around Cordis plugins. The walkthrough covers installation, configuration and tests with local and hosted models. The speaker reports a two-minute install that pulled 530 packages. A dependency warned that it required Node 22.19 while the machine used 22.16, though the app still ran.

The web profile exposes 37 plugin rows, including model access, sessions, permissions and shell tools. The code preset lets the model write a TypeScript program against a generated SDK to combine tool actions. The speaker explains this approach but leaves its performance test for a later video.

For the local LLM test, a configuration overlay points the harness at Ollama on localhost. The setup initially rejects the missing API key; a dummy key lets it proceed. The local qwen3:4b model takes 276 seconds and produces no FizzBuzz file. The speaker says it runs on CPU, while a larger model of roughly 17 GB cannot load on the machine with an 8 GB GPU and about 15 GB of system RAM. A hosted model completes the same task in 13 seconds. These results describe the demonstrated runs, not a general limit on local models.

The web UI runs on port 3080 and shows file changes and script output. The review also notes release-candidate status, some Chinese preset labels and no published DeepSeek score on the cited benchmark. The speaker cautions that the quoted Claude Code and Muse Code scores came from Meta's own evaluation setup.