Interesting that they went NoPE everywhere — everyone else hedges with RoPE in the local layers. Feels like the linear-attention stuff (Kimi Delta) is quietly doing the positional work so they can get away with it. Curious to see if it holds up at frontier scale.
Great breakdown. After using Kimi extensively, it's fascinating to see how architectural choices like KDA and NoPE translate into such strong real-world performance. Really impressive engineering.
"Interestingly, Kimi K3 got rid of all RoPE layers and uses NoPE (No Positional Embeddings) everywhere instead."
It just baffles me that this even works at all. Doesn't it just become a token soup? Is attention that precise that a second token can tell its the second token just because it learns to accumulate something in the embedding space without any sort of inductive bias?
When you have recurrent blocks in your model, you implicitly have a timestep T(amount of recurrent steps). Similar to Diffusion Transformers, it then becomes valuable to encode the knowledge of where you are in this chain somehow. NoPE is more flexible than RoPE for this.
So, unlike what leaders of western labs labs would like you to believe (that Kimi is just the result of distillation attacks), they are introducing new and novel approaches.
Even if they are distilling, I don't particularly care. Anthropic and others have been distilling copyrighted material by to build these models, largely without permission.
"Kimi is largely a byproduct of distillation" and "Kimi is introducing new and novel approaches" are not mutually exclusive, and I'm not sure it's clear from the paper how much of the improvement comes from the new approaches. So I wouldn't take the new approaches to be much evidence about whether the distillation attacks occurred.
Moreover, Distillation is a misused term here. Distillation means training a smaller student model using a larger teacher model to completely mimic its behavior. As in you take a base model and create its smaller “turbo” version. Like distillation in Chemistry it means it’s %100 purified version of its teacher.
Just tried K3 out for the first time today and it's a legitimate threat.
Temporarily (maybe permanently) using it as my daily driver but it's wild how comparable it is to Opus 4.7/4.8 (what's been my go to for a bit now—wrote a quick post on what I found today [1]).
Genuine question: how reproducible / usable / verifiable are these architectures from the published documentation? Are they similar to PDF/DWG/PSD specifications, where the format look like an open spec at first sight until you attempt to implement it and realize the crucial implementation details are undocumented?
I feel like the Kimi team is amongst the best in the industry to pick and choose what is meaningful from the other models. For example, avoiding the expensive and empirically uncertain mHC in favor of simpler residuals. Latent MoE.
My only doubts are around Linear Attention instead of DSA as this is inherently lossy. You are kind of banking on that your query is inherently in the embedding space of the model already and can be lossy.
Anybody getting the result that Kimi 3 is more expensive than Opus 5 or Sol on Cursor? Pretty sure Kimi 3 sucked up a good chunk of my ultimate plan in a few prompts. Anyone have any tools or ways to understand per model usage towards cursor subscriptions? I know there are alternatives to cursor just haven’t made the move yet. (Edit spelling)
kimi K3 felt on par with my claude opus 4.8 in my first tests and is exceptionally good at frontend and website motion design, but I still find it hard to switch back and forth between claude and kimi, and i’m not sure if others are experiencing the same thing.
They may have all this amazing architecture but Kimi has been super dumb recently. I reckon they’re under compute pressure and quantising to stay afloat. I was a heavy k2.5/2.6 user earlier in the year and built serious features with it, but even k3 now does stupid shit like fail tool calls and get stuck in endless thought trains. K2.6 was spinning its wheels on a problem for over 10 minutes today and then deepseek v4 fixed it in under a minute.
I wonder if Kimis KDA architecture is a way closer to online models(models that learn during interference).As KDA updates its state matrix:
St =St−1 +βt (vt −St−1 kt )ktT
The expression (vt −St−1 kt )ktT is literally the gradient step of a linear regression model updating its weights (S) on a single sample using Mean Squared Error loss.
Interesting what other labs comes up with from this.
I know some of these words, but I'd like to understand all of it better. Would anyone be so kind as to recommend a good learning resource for LLM design and the math involved?
31 comments
[ 0.22 ms ] story [ 22.6 ms ] threadIt just baffles me that this even works at all. Doesn't it just become a token soup? Is attention that precise that a second token can tell its the second token just because it learns to accumulate something in the embedding space without any sort of inductive bias?
Temporarily (maybe permanently) using it as my daily driver but it's wild how comparable it is to Opus 4.7/4.8 (what's been my go to for a bit now—wrote a quick post on what I found today [1]).
[1] https://graybearding.bearblog.dev/kimi-k3-is-insane/
My only doubts are around Linear Attention instead of DSA as this is inherently lossy. You are kind of banking on that your query is inherently in the embedding space of the model already and can be lossy.
Something is wrong at moonshot.
The expression (vt −St−1 kt )ktT is literally the gradient step of a linear regression model updating its weights (S) on a single sample using Mean Squared Error loss.
Interesting what other labs comes up with from this.