SoloQ's calibration-free quantizer cuts diffusion-LLM inference costs by 45%

Startup SoloQ released an open-source post-training quantizer for diffusion language models that needs no calibration data. The 4-bit version reportedly cuts FLOPs by 45% and KV-cache size by 58–60% with under 0.5% perplexity drift.

• FLOPs per token drop from 1.8 to 0.99 TFLOPs, KV-cache from 12 GB to 5 GB
• Perplexity rises from 7.12 to 7.15, BLEU falls from 27.8 to 27.5
• A100 latency drops from 312 ms to 186 ms, cost from $0.35 to $0.19 per 1M tokens
• Plugs into Hugging Face Diffusers with one command, no calibration set needed