Qwen3.6-35B-A3B sparks 3090 upgrade debate

The official benchmarks suggest Qwen3.6-35B-A3B is a capabilities bump, but not a clean intelligence win over Qwen3.5-27B. My read: this is an efficiency-and-tooling upgrade first, and a raw general-knowledge upgrade second.

• –On the Hugging Face card, Qwen3.6-35B-A3B is a 35B total / 3B active MoE with native vision support, tool use guidance, and 262K native context, so it is clearly aimed at agentic workflows.
• –The benchmark table shows it is competitive with Qwen3.5-27B rather than obviously dominant on broad knowledge, while looking stronger in several agent and vision tasks. That matches the MoE pitch: specialized throughput, not a simple dense-model leap.
• –For a 3090, the main risk is not just model weights but total VRAM headroom once llama.cpp, ComfyUI, Whisper, and KV cache all compete at once. The user’s concern about spikes is valid.
• –RAM offload is possible in principle, but it is a fallback, not a free lunch. It will usually preserve functionality at the cost of latency and, in the worst case, responsiveness under tool-heavy workloads.
• –The post is useful because it asks the right question: for local users, the deciding factor is often not benchmark rank but whether the model stays stable under real concurrent GPU load.