Qwen3.6-35B-A3B hits ~190K context on 8GB VRAM

Hot take: this is less a model announcement than a useful real-world stress test showing how far sparse MoE inference can be pushed on consumer hardware when the memory layout is tuned carefully.

• –The main value is the configuration recipe, not just the raw benchmark numbers, because it shows what actually moved throughput at very large context.
• –The post suggests Q5 materially outperforms Q4 for long-context reasoning on this model family, which is a useful signal for anyone optimizing quality vs speed.
• –TurboQuant KV cache appears to be the key enabler at ~190K context, making the setup much more viable than standard cache behavior.
• –The Linux + DDR5 emphasis is believable and practical: bandwidth, paging behavior, and mmap/mlock choices likely matter more than people expect.
• –`n-cpu-moe` tuning is the most interesting knob here, but the author is still in the exploratory phase rather than presenting a universally optimal value.