← sources

50% lower error than Strata 3XS, 27% lower error than UD v3.1 3XXS, with 120+ TPS/ 12% faster, on a 30090 + 64GB

source: reddit (reddit:LocalLLaMA)published 2026-11-06 20:16ingested 2026-11-07 02:42

retained text

Hello everyone!

Over the past few weeks since Strata came out, I had many requests to add Qwen Flash Next to LlamAmpere. It was some very impressive engineering, but the models originally shipped (3XS and 2XS) showed very poor quality (.17+ KLD for just the 3XS variant, which uses a far lighter compression formula than more common sources like Unsloth). My initial work had llamAmpere running 20% faster and \~15% lower error, but testing was showing inferior results vs 4bpw Qwen 27b, so I did not find it worth releasing.

I continued to make progress on the AmpereInteger+ quantization format (I will release an encoder soon), and Unsloth released their v3.1 which provided a great new base.

Happily,[ LlamAmpere v0.6.3](https://github.com/JakeATX/llamAmpere) is now available with a new model and quant updates that show significant speed and quality gains over other options.

At a depth of 100K, QFN runs at \~120+ TPS, with prefill >3000/s through the first 100K. It is a 50% error reduction vs the version of QFN available for this card size a few days ago, and 27% vs UD's improved quant.

https://preview.redd.it/wcumk817v10i1.png?width=1108&format=png&auto=webp&s=4c321fbbbf328cd95163ecc51a7d6587fa1308c7

I do have a 3XS class file that runs \~10% faster than the version above, but as it is only \~15% better quality than the original Strata 3bpw model, I don't recommend it.

This will also work well for 12+64GB configs. As with Strata, your performance will depend on the speed of your CPU and RAM as well. I do not recommend any quant of QFN for configs under 54GB of total memory, but you can always try the coder pruned variants as well.

Enjoy!

Model: [https://huggingface.co/jakeatx/qwen3.8-flash-next-Swift1.5-ATX-AI3](https://huggingface.co/jakeatx/qwen3.8-flash-next-Swift1.5-ATX-AI3)

LlamAmpere: [https://github.com/JakeATX/llamAmpere](https://github.com/JakeATX/llamAmpere)

document fetched 2026-11-07 00:40 · canonical https://reddit.com/r/LocalLLaMA/comments/1xpmzg7/50_lower_error_than_strata_3xs_27_lower_error