← sources

UPDATE - New Quants: Swift 1.5 Flash Next - Heretic v2 - ReverseEngineer

source: reddit (reddit:LocalLLaMA)published 2026-10-30 19:02ingested 2026-10-31 02:40

retained text

Original post: [https://www.reddit.com/r/LocalLLaMA/comments/1xix35p/release\_swift\_15\_flash\_next\_heretic\_v2/](https://www.reddit.com/r/LocalLLaMA/comments/1xix35p/release_swift_15_flash_next_heretic_v2/)

Hello! At request I've released a few more formats of my Swift Flash Next Heretic reverse engineering edition.

The original model is the IQ4\_NL, the first quant I made and the best performance/size ratio. That's my favorite but its hefty and I appreciate people wanting other options.

A: IQ4\_XS - Based on the Unsloth IQ4\_xS recipe, with about 1gb overhead added where it was worth the performance to size tradeoff. Estimated \~9.7% reduction in expert quantization error mass for +0.96 GiB over the plain UD-IQ4\_XS recipe. Quantized with an identical imatrix to the IQ4\_NL quant: 94.7 GB on disk compared to 109gb for the IQ4\_NL.

B: IQ3\_S - most requested. Custom quant recipe based on data gathered from the model, in the same way the other quants were done. High quality and model specific. I modeled mine off of ISTALab's with a few adjustments. 

IQ3\_S uses a more aggressive imatrix distribution than the IQ4 to try to preserve strong reverse engineering capabilities at lower bit widths, since it's getting small. While the IQ4 builds are 70%/30% distribution in the imatrix merge, this one is 80%/20% - better reverse engineering and reasoning ability at the cost of some general knowledge and accuracy. 85gb on disk (ISTA GSQ-RCO IQ3\_S is 83.6 gb, so 1.4gb overhead - it's worth it for the accuracy improvements.

[https://huggingface.co/trubisky/Swift-1.5-Flash-Next-Heretic-ReverseEngineer](https://huggingface.co/trubisky/Swift-1.5-Flash-Next-Heretic-ReverseEngineer)

Cheers.

document fetched 2026-10-31 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xjo7rt/update_new_quants_swift_15_flash_next_heretic_v2