← sources

Gyro quants looks like EXL3 but kinda closed source. please explain why we need this.

source: reddit (reddit:LocalLLaMA)published 2026-10-24 09:17ingested 2026-10-24 11:01

retained text

Recently this new quant has popped up called Gyro. It is gaining popularity because it's being inserted into many different engines, especially the recently popular Strata.

So i took a dive into the code to see the differences between Gyro and EXL3, to me it looks like simply a different flavor of EXL3 quantization, but it's not interoperable and the recipe to make it is closed source for now.

|Property|Gyro-S / Gyro-M|EXL3|
|:-|:-|:-|
|Quantization family|QTIP / Trellis-coded quantization|QTIP / Trellis-coded quantization|
|Hadamard rotations|Yes|Yes|
|Trellis state size|15 bits|16 bits|
|Encoding unit|128-weight blocks|256-weight tiles (standard path)|
|Codebook|Explicit LUT with 2,048 2D points|Procedural codebook|
|Weight reconstruction|State hashing + lookup table|Procedural value generation|
|Bit packing format|`TQ2_T`, `TQK6`, `TQK7`|EXL3-specific packing|
|Direct interoperability|Not demonstrated|Not demonstrated|

I did not find any published evidence that it's better than EXL3 on either decode speed, or perplexity/KLD/top 1%.

I'm not dunking on Agentionai but I kinda hate how this only makes the community more fragmented without providing any benefit. EXL3 is not very publicized while this Gyro is gaining popularity with aggressive brading (socials and stuff).

"*Agention Precision Rotor*"... please, proof....

document fetched 2026-10-25 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1xe41qe/gyro_quants_looks_like_exl3_but_kinda_closed