各主流LLM本地部署的启动参数配置大全（基于LM Studio）

2026-08-04 更新于2026-08-04 13阅读

本文基于LM Studio，总结当前各主流LLM本地部署的启动参数配置大全，以发挥个人算力机器的价值，文中所有的配置都经过验证可行。

LuffyTheFox/qwen3.6-35b-a3b-uncensored-genesis-hermes-v6

启动参数配置：

类型	参数项	参数值
Model Information
	Model name	LuffyTheFox/qwen3.6-35b-a3b-uncensored-genesis-hermes-v6
	Format	GGUF
	Quantization	Q8_0
	Arch	qwen35moe
	Capabilities	V，T，R
	Domain	llm
	Size on disk	37.80G
Context and Offload
	Context Length	65536
	GPU Offload	40
Advanced
	CPU Thread Pool Size	16
	Evaluation Batch Size	8192
	Physical Batch Size	512
	Max Concurr..	2
	Unified KV Cache	开
	Context Checkpoints	32
	Reasoning Budget Message	空
	RoPE Frequency Base	不选，Auto
	RoPE Frequency Scale	不选，Auto
	Offload KV Cache to GPU Memory	开
	keep Model in Memory	开
	Try mmap()	关
	Seed	不选，Random Seed
	Number of Experts	8
	Number of la...	16
	Speculative Decoding	Off
	Chat Template	默认
	Flash Attention	关
	K Cache Quantization Ty...	不选，Experimental
	V Cache Quantization Ty...	不选，Experimental
Inference	System Prompt	默认
Reasoning	Enable Thinking	开
	Reasoning Budget	不选，Unrestricted
Setting	Temperature	0.1
	Limit Response Length	不选
	Context Overflow	Truncate Middle
	Stop Stri...	默认
Sampling	Top K Sampling	40
	Repeat Penalty	选中，1.1
	Presence Penalty	不选
	Top P Sampling	选中，0.95
Structuerd Output	Structuerd Output	关