Good stuff!
Thanks for giving us something to play with while we wait for GLM 4.6 Air. I humbly request that you finetune 4.6 Air on this dataset too when it comes out. I think your Iceblink V2 is a noticeable improvement over V1, and from my initial testing, it seems like the best GLM finetune available right now.
Great work as usual!
Thank you!
For sure, I was on the fence about waiting for GLM Air 4.6 but got impatient. Assuming they don't make any big architecture changes that's the plan for when it comes out.
As a base I'm 50/50 on it compared to some of the old 70b tunes. It's smarter and has a ton of world knowledge, but I'm not a huge fan of how it writes. I'm also running 2x 3090s and don't really have a favorite model currently.
70B are good but feel a bit dated and sometimes unstable, air is good but feels a bit tame / assistanty, smaller models like mistral are fun but stupid. I'm hoping GLM 4.6 Air comes out / is good currently I think. Or a new dense model with decent world knowledge.
I feel everything you just said, zerofata. Everything is a tradeoff. Here's hoping GLM 4.6 Air is a step forward, and maybe we'll get some gems in 2026.
I also concur with you both.
And @zerofata , I like very much Iceblink v2, quite more than the v1. Last time I was that pleased with a "fast" creative model was when playing with Electra R1 70b and Legion 70b. Iceblink 2 gives me a little feel of the 355b model in terms of prose amount and readability, minus some overall intelligence and coherence. But with the inference speed to enjoy it and regen if necessary. Congrats, I really hope for GLM 4.6 Air to.. air soon! Its 355b version is a step forward compared to 4.5 in terms of RP intelligence, for what I could test myself.