SKB

the cake is a lie

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[SKB @skbzz7]

Open source won.

[SKB @skbzz7]

I was going to write a whole thing about information monopolies and why open source has to win, but I got lazy. Assume I made several excellent points.

[SKB @skbzz7]

Anthropic has had enough PR disasters to teach a full course by now.

[Julian Schrittwieser @Mononofu]

I’m so excited that @JensenHuang is a believer in open source now, looking forward to the CUDA and GPU driver open source release!

[Tibo @thsottiaux]

Unbelievable excited for what’s coming together. Tomorrow is feeling codexy

[SKB @skbzz7]

Noice

[jtydhr88 @jtydhr88]

It’s time ComfyUI had real VFX capabilities.

Real-time preview while you adjust, multiple effects chained together, changes propagating down the chain, proxy loading for large files…

No more switching between workflows, what you see is what you get, and everything happens on a single canvas.

Color grading, curves, LUTs, stylization, green-screen keying, frame interpolation, blur, sharpen…
if you can think of it, it’s here.

All of it, in ComfyTV.

[SKB @skbzz7]

Stay strong, my poor GPU. Your suffering is almost over.

[SKB @skbzz7]

Using your entire weekly usage limit in a single day because the thought of letting it reset unused is unbearable. Reset-Day Syndrome.

[SKB @skbzz7]

@mervenoyann feels good

[SKB @skbzz7]

Claude is better and I’m prepared to fight for this completely independent opinion. Also, entirely unrelated: thank you Anthropic for the free 6 months of 20x Max.

[Vadims @vadi_ms]

This is me talking to my computer without making a sound.

After just a month of collecting data, our model is already approaching dictation in accuracy. We were surprised to see that it generalizes to unseen participants as well!

(1/n)

[SKB @skbzz7]

Stupid question: what if real art starts when we stop writing prettier prompts and start messing with the activations until the model hallucinates on purpose? I don’t mean the hallucination after it becomes an output. I mean the weird internal moment before the model remembers how to behave. I’d write more, but my English ran out.

[Anthropic @AnthropicAI]

New Anthropic research: A global workspace in language models.

Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with.

We found a strikingly similar divide inside Claude.

[SKB @skbzz7]

Oldies. C4D+AE

[SKB @skbzz7]

Different projects. Different communities. Same comment every time. "Why?"
I wish I knew. Making weird little things just feels good.

[SKB @skbzz7]

@creator_kachun Not yet. I've got quite a few nodes waiting to be cleaned up and released.

[SKB @skbzz7]

LoRA testing is pure brain rot. I added one helper, then ten more, and now there’s a Node package generated entirely by my mental illness.

[SKB @skbzz7]

Nothing special, just a small radio player for ComfyUI.
Supports YouTube/live links and direct radio streams. That’s it.

https://t.co/G2kSSFmfst

[SKB @skbzz7]

Dropped GreyZone. For Krea 2.
Just that strange early 3D atmosphere.
Not clean, not modern, not pretty in the normal way. Just old 3D mood and weird places.

https://t.co/7gsD7A2d5h

[SKB @skbzz7]

Also, here’s a little sneak peek at the next LoRa I’m working on.

[SKB @skbzz7]

@icantevenfilms It’s more of a sequencing/testing toolkit for automating the settings you’d normally have to change manually every run.

[SKB @skbzz7]

@hellorob

[ilker @ailker]

🎨 I trained 999 Style LoRAs with the @fal Krea 2 Trainer.

Every single image in this video is a different LoRA.

That's 999 unique Style LoRAs, each trained for just 100 steps at 0.00035 LR, taking an average of ~30 seconds to train.

As I said before, @krea_ai 2 continues to prove that it's an incredible model for open-source Style LoRAs. Next week I'll be sharing even more LoRAs, along with a few surprises.