Qwen3.6-35B-A3B hits 30 t/s at 128K

The punchline is that long-context local coding is no longer a “maybe someday” story on a single consumer GPU, but a tuning problem with sharp cliffs. Once the author found the right MoE offload balance and KV layout, Qwen3.6-35B-A3B became fast enough to keep a Claude Code-style workflow practical.

• –The key benchmark is the depth curve: the model stays strong as context grows, instead of collapsing the way the dense 27B setup did.
• –The post makes a convincing case that file size and offload balance beat “better” quant labels if they let more experts stay on GPU.
• –The eval lesson matters: self-written tests overstated quality differences until the author switched to one shared harness and deterministic sampling.
• –The hybrid KV finding is the most interesting systems result here: compressing everything was slower than selectively promoting hot layers to q8_0 at long context.
• –For local agent users, the real threshold is not peak tokens/sec at empty context; it is whether the model stays responsive at the conversation depth you actually work in.