Comment by tough

We use GLM at work and YMMV a lot depending on the provider. So i guess some have bugs on their inference pipeline, or serve too agressive quants