A real test of cutting Claude token use with a local model
A company is testing a workstation with an to run and reduce part of its Claude token use. The setup runs Qwen3.6 27B MTP Q8_K_XL with llama.cpp on Windows 11. In coding work, the model feels somewhat close to Claude Sonnet, but weaker and slower.
It works better than Claude Haiku for this use case, while Claude Opus remains far ahead. For its relatively small size, the model is surprisingly good at reasoning and tool calling. Its main weakness is missing knowledge.
Giving it tools such as Context7 and Serper, so it can check documentation and search the web, made it less likely to invent class names, field names, , and similar details. The major problem is stability: during in VS Code with the Copilot extension, the agent sometimes stops at random.
Key points
- The goal is to reduce some Claude token use by running a local model on company hardware.
- Qwen3.6 27B MTP Q8_K_XL feels better than Claude Haiku, but weaker and slower than Claude Sonnet.
- Claude Opus is still seen as much stronger.
- Context7 and Serper helped the model check documentation and web results instead of guessing coding details.
- The main blocker is random stopping during VS Code coding-agent sessions with Copilot.