#ai #llm #selfhosted I've found myself using a stack of weird tricks to get reasonably reliable performance out of local...

#ai #llm #selfhosted I've found myself using a stack of weird tricks to get reasonably reliable performance out of local models.First and foremost, I use a revised chat template for Qwen 3.8 Flash Next that fixes a few bugs as well as gives it more examples for batched tool calling, otherwise it tends to do one at a time.Second, I use a pi extension called pi-vcc, which is deterministic compaction. Super fast, no LLM involved.Third, because pi-vcc can be a little too aggressive sometimes, I have a skill that injects whatever task I've give a subagent directly into the system prompt so it can't get compacted away. This one change has probably increased acceptance rate more than anything else.