llama.cpp update breaks older GGUF chat templates in coding-agent workflows
computehungry · reddit · 2026-07-28
A llama.cpp update broke preservethinking behavior in older GGUF chat templates, making the model much dumber in coding-agent workflows.
- Recent commits changed how older templates behave.
- The breakage affects agent-style use, not just casual chatting.
- Adding kwargs does not fix it.
- The suggested workaround is to explicitly point llama.cpp at the DeepSeek-V4 chat template:
--chat-template-file /path/to/llama.cpp/models/templates/deepseek-ai-DeepSeek-V4.jinja
More from coding & agent
- FindDomain lets AI agents check and register domains through an API — geek_guy1 · 2026-07-28
- A talk on AI engineering argues evals, memory, and tools matter more than scattered tool choice — Pavan_Belagatti · 2026-07-28
- Well MCP connector plugs AI into invoices, companies, and contacts — modelcontextprotocol · 2026-07-28
- Wikipedia MCP server gives LLMs multilingual search and page fetch tools — modelcontextprotocol · 2026-07-28
- Agent builder adds four messaging providers and still can’t reach users reliably — iamstanty · 2026-07-28
- Prompt scaffolding for Claude agents turns into technical debt as models change — TechPreacher · 2026-07-28