Weird but interesting
It clearly have Qwen thinking style. It mostly write from first person essentially simulating "person that think", not "thinking". Which is big plus.
But it clearly NOT Qwen. A bit more messy and actually more fun in creative writing.
Have abysmal Russian and actually prone to looping when interacting with it.
Well, it seems to be Qwen with another label. 35B model out of nowhere, huh?
No it have weird architecture. Weird enough to require it's own llamacpp fork. I thought same thing but the fuck is MBSA? The hell is Manifold-Constrained Hyper-Connections?! No, it's clearly not another Qwen. It's look similar but only on surface. Even vocab/tokenizer are different. That is actually why it is abysmal in Russian.
Okay, maybe it has vibe-coded architecture and distilled on local Qwen model? Distillation was not including many Russian prompts
Okay, maybe it has vibe-coded architecture and distilled on local Qwen model? Distillation was not including many Russian prompts
everyone's a critic
everyone's a critic
You don't need to be a cook...
Thanks for the thoughtful feedback! Just to clarify, Maxwell-V1-Flash was pretrained from scratch, using our own architecture and training data
We’d love for you to keep trying it out and share any further feedback—it would be very helpful to us.