Oct DeepSeek Usage Tips

Using the DeepSeek Harness is recommended first. Several users report tasks finishing almost in one pass, but some feel Pi is faster and more accurate.

Main model setup: if the budget allows, use the official API. For better value, choose the Workbuddy plan, then pair it with Opus for planning and GPT for review.

The OpenCode Go plan offers generous quota, but some users tested it and found timeouts during peak hours and quality below the official API. When choosing a plan, check speed and stability.

For writing code, it's recommended to enable the Max thinking level, which thinks longer but works better. One user spent 28 yuan running through ten cases, and with a skill configured, the results were good.

Using Pi with the high level, all ten real programming tasks passed, at under 5 US cents per task, nearly 3 times cheaper than Opus 5.5 low.

Claude Code and Codex can call DeepSeek just by changing the Base URL, and you can also run it directly inside Codex with magpie.

The Shanghai AI Laboratory platform gives you quota when you check in daily, and after student verification you can use domestic models such as DeepSeek free for one year.

For local deployment, choose the quantized version by memory: use 1-bit with 128G of memory, 3-bit with 256G, and 64GB of memory can run Unsloth's GGUF.