Output speed on a model that can actually do agentic coding.
Target: 1,000 tok/s per stream on a model that can agentic-code, with only 10–30B active parameters — the size that still runs on hardware you can rent by the hour.
Today, measured on the fleet: 500 tok/s FP4 decode (Qwen 3.6 35BA3B NVFP4) for 1.5$/h, 150 tok/s (GLM 5.3 Flash NVFP4 Dflash) for 6$/h, 500 (Qwen 3.8 27B NVFP4) for 7$/h · aexol.ai/models