← sources

ijah --- M5+ Mac Engine for Qwen 3.8 Flash Next / Swift 1.5 Flash Next. Optimal Quality and Speed on 128gb

source: reddit (reddit:LocalLLaMA)published 2026-10-19 17:22ingested 2026-10-20 02:39

retained text

Sorry, yet another flipping engine!

A few weeks ago, I made a speed focused experimental fork of Splash for M5 Max users recently called [Splish](https://github.com/publicExcess/splish). One of the most requested features was Qwen 3.8 Flash-Next support.

Honestly, my experience with Flash-Next was underwhelming before this. 27B is such a beast that I never saw margins from Q4 versions big enough to justify the RAM footprint.

Then came [Sushi](https://github.com/beamivalice/sushi). What a slept on engine. Totally flipped my opinion on FN.  With EXL3 trellis weights, you can fit on [5b -6/8b models](https://huggingface.co/SiliconSpecies) and they cook.

Inspired, and borrowing heavily, I wrote new kernels to see what else was left on the table. The result is [Ijah](https://github.com/publicExcess/ijah).

**\*\*\***

**With 5b-6/8b Model:**

***Quality***\*:\*  About 4–5× closer to the bf16 model than Jundot's oQ4e (KLD 0.033/0.009/0.015 vs 0.128/0.036/0.078), \~2× closer than Sushi-4bpw.  
***Decode***\*:\* \~100 tok/s (90–110 through the server) —  \~6% behind oMLX but with 5-bit experts. With Inco's DFlash2  drafter released today I'm getting  \~120 tok/s. Will charts later..

\*\*\*

So it'll lose in a drag, but honestly I think it more than makes up for it. Please give it ago and let me know how it goes?

Support my token addiction  [ko-fi.com/severalviolins](https://ko-fi.com/severalviolins).

document fetched 2026-10-20 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1xa30fc/ijah_m5_mac_engine_for_qwen_38_flash_next_swift