Qwisp: MoE expert-streaming engine runs Qwen3.6-35B-A3B on 8GB Macs

Qwisp is a new open-source local inference engine built specifically for one model — Qwen3.6-35B-A3B (MoE) — on Apple Silicon. By narrowing its scope to a single target, it delivers three capabilities general-purpose engines typically can't: a strict bit-exact lossless mode where the token stream matches greedy decoding exactly, MoE expert streaming that keeps only the active parameter slice resident in RAM while streaming the rest from flash so the 35B model fits on 8GB Macs, and a raw-Metal decode core that bypasses the MLX op-graph for hand-issued command buffers. The developer reports 85-91 tok/s in strict-lossless mode on an M1 Max/64GB, with RAM-forced approximations for lower-memory tiers. The project is now crowdsourcing a deterministic benchtest across the community to map performance across chip, RAM, and SSD combinations, particularly for under-tested 8GB/16GB/24GB configurations. For engineers, it's a concrete example of how model-specific, hardware-level optimization — raw Metal buffers, speculative decoding, expert streaming — can outperform general-purpose inference frameworks on constrained hardware.