Key Info
Ahmad Awais (CEO of Command Code) shared an engineering deep dive arguing that the common belief "open models are bad at tool calling" is almost always a harness problem rather than a model problem. The analysis is based on examining billions of tokens processed with DeepSeek inside Command Code's open source AI CLI.
Highlights
- The write-up challenges the assumption that open models are inherently weak at tool calling.
- It frames tool-calling failures as a harness (scaffolding) issue, not a model capability issue.
- The findings come from analyzing billions of tokens run through DeepSeek in Command Code's open source CLI.
- Awais also recommends Command Code with tool call and cache repairs for users working with DeepSeek.
- A link to the full engineering deep dive is provided for further reading.