Kimi K3 Architecture Overview and Notes(sebastianraschka.com) |
Kimi K3 Architecture Overview and Notes(sebastianraschka.com) |
(As an aside, this is why the "open weights are not open source" thing is a complete misunderstanding. The weights themselves along with the documentation give you enough to fine tune the LLM. You can't rebuild it from scratch, but you can't do this even with the data anyway (because of randomness!))
They could give you the random seeds? (Assuming you carefully train in such a way to remove other sources of randomness, like concurrent execution.)
> You can't rebuild it from scratch
There's is extremely clear and misunderstanding-free.
Open weights is not open source.
> It's entirely reproducible from the available documentation
You have a very interesting understanding of "reproducibility", I'll give you that :)
But even with that, there are plenty of technical details (especially in regards to the training process) missing from the tech report that leads to these weights not being reproducible in any sense of that word.
Reproducing the exact training run, however, is basically impossible without the original dataset and training pipeline (here meaning all of the code + infra involved in actually executing the pre and post training loops). Also, it would be exorbitantly expensive to do if you weren't also a lab trying to train a similar model.
But you can still scale the architecture down and experiment as a solo researcher using the published research. There are probably some open source implementations already on GitHub for any given big open model release.
Transformers are very "mendable" in that you can permute the architecture in crazy or random ways, and still basically always end up with get a coherent LLM. The difference comes down to training efficiency, inference efficiency, and usually minor differences in performance.
Hyperparams and stuff, I mean it's standard to do a sweep anyway.
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?
And yes accumulation is a good intuition for what's going on. You could imagine a part of the attention head that just kept writing to the same part of the residual stream causing that to keep accumulating (simply via attention summation) as more input tokens come in thereby functioning as a kind of index without the need for any positional encoding.
"Kimi K3 uses no explicit positional embedding (NoPE), and instead encodes positional information implicitly through the recurrent gating and decay mechanism of KDA."
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.
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.
Source: my bank balance
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]).
Plus I can actually see the thinking traces and I’m not getting downgraded/denied by anthropic constsntly…
that's a weird way of describing a near frontier open weights un-crippled useful coding buddy
do you work for OpenAI or Anthropic per chance?
And yes, it's a legitimate threat to frontier models. Both K3 and the distillation approach.
Like the distillation “attacks” Anthropic conducted against literally millions and millions of different creators?
There's been some distillation.
Just like how Elon said in court they distill to make Grok.
Something is wrong at moonshot.
Because it's RNN-like, it has an inherent idea that X comes before Y which comes before Z in the sequence XYZ. Transformers, by default, don't have that. They operate on sets, unordered collections of unique items. They have no idea where those items are in relation to each-other so you have to clue them in.
Because There are 3 KDA layers per attention layer, and 3 KDA layers before the first attention layer, every single token position is going to be able to learn information about where it is in the sequence before the first actual attention layer.
RoPE is actually a bit destructive, so being able to omit it like this is very convenient. Models like Gemma-4 have a similar structure with 5:1 Sliding Window Attention (SWA) layers for every global attention layer. These are cheap, shitty attention layers which handle local information and which go in-between the big powerful ones, KDA serves the same role in this model. In Gemma only the SWA layers have RoPE while the global attention layers omit it. SWA is actual attention, even if it only operates on a small sliding window, so it needs the positional embedding. KDA isn't, so it doesn't.
>Curious to see if it holds up at frontier scale.
I don't know how much more frontier scale you can get than this, but yes, there's no reason why that wouldn't work at larger scales. Honestly, more parameters just makes it easier for the KDA layers to communicate that positional information better.
> Seems to benchmark
yes, but in human usage the differences show up
SWA has a similar issue. Unless you keep the entire KV prefix lying around (which is not unreasonable: you retain flop + bandwidth benefits but lose capacity benefits), you need to start 1 window back from the rollback point, in order to refill the sliding window before going into normal prefill.
If you want to train it from scratch you need data, yes. But presumably if you are doing that there is a reason you want to do it.
You lose nothing without access to the original data - you can do every single modification without it.
That is unlike open source where you (mostly) need to source code to modify it beyond what the original designed originally thought.
Have you ever tried reproducing even a small neural network exactly if you train on GPUs on more than one machine? I have and it is pretty close to impossible, and I'd argue actually impossible at scale.
Their 200 USD tier is pretty good and I might move to them because their annual discounts also make it even cheaper. I dislike the way how Opus/Fable write and Kimi has less of that slop writing problem, but also ends up feeling kinda slower (especially on Max reasoning), technical abilities are not 100% the same but close (better in some tasks, worse on others).
Edit: well, crap.
Model overloaded retrying (6/10) 1m 40s
Though in principle saving random seeds is a lot less hassle than keeping entire checkpoints around: your random seeds would fit on a floppy disk or even a tweet. The checkpoint is basically as big as the model.
1. https://news.ycombinator.com/item?id=49019012 (original chinese)
2. https://news.ycombinator.com/item?id=49052912 (translated english [pdf])
Though in practice you can get all the benefits of both determinism and (that kind of) randomisation by using a PRNG and saving the seed you are using.
It's an open question roughly on par with P vs NP whether true randomisation is ever necessary, or whether PRNGs are enough. So far we haven't found any problem or algorithm where true RNG is necessary and good PRNG ain't enough.
He was talking about two different things, hence the parentheses. The architecture is reproducible, not the model weights.
But what's the point of even saying that? Of course it is, otherwise how is it supposed to run in the runtimes? You cannot release model weights that others can run, without also releasing the model architecture, it's in the code at the very least...
The K3 arch can be implemented from the spec.
A PDF reader can not be implemented from the spec.
I share this expectation. But this is an interesting empiric question that deserves study; even if just to confirm what 'everyone knows'.
Heck even ffmpeg introduce randomness when stitching together downloaded chunks from youtube. By design.
I don't see the connection? Most local builds are done with debugging on and optimisation turned off anyway.
And what you deliver to your customers is usually something you produce on your CI/CD server, not what's on any developer's machine.
(And if you want reproducible builds https://en.wikipedia.org/wiki/Reproducible_builds you can't optimise for a specific wall clock time.)
> Also, most (optimizing) compilers ”optimize” the code for a fixed amount of time, leading to better optimized binaries on faster computers.
That’s why you should give your developers computers that have slow clocks!