Qwen3.6-35B-A3B sparks 3090 tuning hunt

Hot take: this is the kind of release that matters less on paper than in the hands of local-LLM tinkerers, because the real product is the performance envelope you can actually sustain on consumer hardware.

• –The model is already being treated as a local inference target, which is a good sign for adoption among power users who care about latency, not just benchmark headlines.
• –llama.cpp tuning now matters as much as model choice: context size, KV cache quantization, GPU offload, and batch sizing will decide whether a 3090 feels usable or cramped.
• –The thread’s low comment count suggests this is still early, with most of the useful signal likely coming from hands-on experimentation rather than consensus best practices.
• –If Qwen3.6 really improves agentic coding, then local users will optimize for stable interactive throughput, since coding workflows punish stalls more than raw single-prompt speed.