Qwen3.6-35B-A3B vs Presto Voice

What users praise

• • Runs 50-90 tok/s on consumer hardware like M1 Pro and RTX 3090.
• • Apache 2.0 license permits commercial use, modification, and redistribution.
• • Strong agentic coding and tool calling capabilities praised by the community.
• • Multimodal reasoning often comparable to much larger dense models like Claude Opus.

What frustrates them

• • MoE architecture may be less accurate than dense 27B for deep reasoning.
• • Quantization quality is critical—poor quants degrade output noticeably.
• • Vision encoder required separately for multimodal tasks.
• • Low-end GPUs (e.g., GTX 1060) achieve only 11 tok/s.

What users praise

• • Fifteen-plus years in restaurant automation gives Presto real QSR operational experience
• • Handles POS and headset provider integration itself, avoiding a lane shutdown at install
• • National rollouts at Wienerschnitzel, Taco John's, and Dairy Queen validate enterprise scale
• • Spectrum-of-models approach targets store-by-store variation in menus, accents, and ambient noise

What frustrates them

• • No independent operator reviews exist in the public data to validate the 95% claim
• • Vendor-published metrics lack third-party audited baselines or methodology
• • Only Toast is named as an integration — other POS stacks are unproven
• • Pricing is undisclosed, making per-lane ROI modeling impossible up front