What's your tps on 3090 + Qwen 3.6 27B in real tasks?

via

LocalLlama

I struggle to wrap my head around all this. My goal is local agent to solve low complexity tasks, in the same harness where I would use frontier models. So naturally this means a large context window, because low complexity can mean a simple-ish fix in a large codebase, rather than just generating some nonsense from zero. So initially I went for Tom's turboquant plus fork of llama.cpp (I'm on Windows) with Qwen 3.6 Q4 and IQ4 models and 200k context window. Well it worked, it can read the entire

More from LocalLlama

Bruh

LocalLlama·2

Qwen3.6-27B-NVFP4 - images

LocalLlama·5

Hybrid on-device inference on Android: llama.cpp + LiteRT + NPU/GPU routing

LocalLlama·5

MiniMax M2.7 AWQ-4bit on 2x Spark vs 2x RTX 6000 96GB - performance and energy efficiency

LocalLlama·3

Qwen3.6-27B at 72 tok/s on RTX 3090 on Windows using native vLLM (no WSL, no Docker), portable launcher and installer

LocalLlama·6

A Dark-Money Campaign Is Paying Influencers to Frame Chinese AI as a Threat

LocalLlama·6