netrunner

Cofounder & engineer.
Running local AI, building things with it, and sharing what I learn.

[netrunner @plotarmordev]

Drama 3 went on a 4bit diet.

My unofficial NVFP4 quant of @fishaudio's Drama 3 drops 28% of its experts and peaks at 19.35 GiB instead of 36.95, with no measurable quality loss in my tests on one DGX Spark.

Capri, my local assistant, explains. Sound on 🔊

Links in the reply:

[netrunner @plotarmordev]

@boardyai @Microsoft @NVIDIARTXSpark Has to be heavily quantized, let me see what I can do, it reserves a lot of RAM for the OS

[netrunner @plotarmordev]

Surface Laptop Ultra by @Microsoft with the new @NVIDIARTXSpark inside and 64GB of RAM.

Purely hypothetically, if one landed on my desk, what would you want me to test first?

PS: it did! That's a pic I just took. Not mine to keep, but I get to put it through some tests!

[netrunner @plotarmordev]

@WasamiKirua @blocks Not dropped yet

[netrunner @plotarmordev]

Dessa by @blocks is crazy to me. I don't think it even belongs under text-to-speech, it feels like a whole new category of model.

It never saw a single line of text. It was trained from scratch only on recordings of people talking, and it listens and talks at the same time. And most importantly, it's small enough to fit in a Macbook!

[Block @blocks]

Meet Dessa, our speech model trained from scratch on human conversation.

Hear it for yourself: https://t.co/FDmoXNPpIh

[netrunner @plotarmordev]

@devopsfortunato @Microsoft @NVIDIARTXSpark Haha let me see if I get enough time with it to do that, I think it has a limit on the ram you can “dedicate to vram” it wants you to leave enough for the OS to operate. So it’s really a personal computer for hybrid ai setup

[netrunner @plotarmordev]

@niklaslenz_ai @Microsoft @NVIDIARTXSpark Will try that! I could only use it a bit today, but it seemed snappy and with a great battery, and extremely silent. Didn’t use it for inference yet.

[netrunner @plotarmordev]

@FutureTechie @FishAudio @NVIDIAAI @MiaAI_lab @ashxhart Exactly, are you going to try it? I have it at hand but I haven’t had time to set it up since it’s a personal computer and on those I don’t like agents moving around 😂

[netrunner @plotarmordev]

Drama 3 went on a 4bit diet.

My unofficial NVFP4 quant of @fishaudio's Drama 3 drops 28% of its experts and peaks at 19.35 GiB instead of 36.95, with no measurable quality loss in my tests on one DGX Spark.

Capri, my local assistant, explains. Sound on 🔊

Links in the reply:

[netrunner @plotarmordev]

Claude is going to change without @trq212.

More than a year on it, watching it grow way bigger than even he expected, and a lot of that happened right here on X with us pushing them.

Can't wait to see what's next!

[Thariq @trq212]

I’ve left Anthropic after more than a year of working on Claude Code and seeing it scale larger than I ever imagined.

I have really loved working at Ant and believe deeply in the mission, and most of all, the idea of a race to the top.

The idea that if we work hard and responsibly, we can raise the bar and inspire others to great outcomes. And if we all push each other to be better, things can be radically good. I will carry the spirit of this with me forever.

My colleagues at Anthropic are the most talented, thoughtful, competent, and earnest people I have ever met. I am glad they are at the wheel and trust them deeply.

To the community on X, thank you for everything. You’ve made this the most fun (and sometimes stressful) job I’ve ever had.

For me, it is time for a bit of a break before diving into something new.

If you want more excellent Claude Code content, you should follow @addyosmani , @lydiahallie, @_catwu, @ClaudeDevs and of course @bcherny who started it all.

[netrunner @plotarmordev]

@u1tra_instinct @yume_arasaki @FishAudio @NVIDIAAI @MiaAI_lab @ashxhart Yep easily fits a 5090 too, has to be super fast there

[netrunner @plotarmordev]

@jurlycat agree it still takes work, but most of that work used to be guessing. now you can ship the rough version, see who actually uses it, and prompt the next one

[netrunner @plotarmordev]

We're a few prompts away from success.

Never in history could anyone say that and mean it. Building something used to take years of learning, a team, or money.

Now the distance between an idea and a working thing is a few prompts and the stubbornness to keep sending them.

Keep prompting!

[netrunner @plotarmordev]

@yume_arasaki @FishAudio @NVIDIAAI @MiaAI_lab @ashxhart 💯 you got it haha

[netrunner @plotarmordev]

@yiz_be_building @blocks Just a way to articulate, that’s what I thought tho 🙏🫡

[netrunner @plotarmordev]

@yiz_be_building @blocks I think I said “ it doesn’t belong to text-to-speech class of models, and it’s a new category 🫡

[netrunner @plotarmordev]

@maxmilneaus @blocks That’s what they call it? I was about to call it that and innovate with my naming scheme 😂

[netrunner @plotarmordev]

@MrV_777 @blocks Not yet, couldn’t find any way to use it

[netrunner @plotarmordev]

@ARomeoSierra @blocks Yep new category of models

[netrunner @plotarmordev]

@whosamberella @FishAudio @NVIDIAAI @MiaAI_lab @ashxhart Thank you! 🙏

[netrunner @plotarmordev]

@kkiran @blocks Yea maybe but if it runs on a MacBook do you really want to run it on an iPhone? Wasted battery etc? You can easily just give it a call

[netrunner @plotarmordev]

👀🙏 thank you @jack!

I guess the only question I have now is, when can we use it? Especially locally

[netrunner @plotarmordev]

@ashxhart @Alibaba_Qwen @QwenDevs @petruspennanen @volatilemarkts @MiaAI_lab That’s a big gap. Is the kernel doing some kind of speculative/draft optimizations?