Ash Hart

AI enthusiast

[Ash Hart @ashxhart]

@jundotkim Great work!

[Ash Hart @ashxhart]

@runsonai Same setup as me, I need another spark though which I’m planning to buy.

[Ash Hart @ashxhart]

Continuing to push @deepseek_ai's V4 hard on the M3U Studio.

There is more speed on the table, and I will continue to update the repo.

https://t.co/GR2MPbEhqS

[Ash Hart @ashxhart]

@0xSero I'll test it now.

[Ash Hart @ashxhart]

Pi is goat.

[Mia @MiaAI_lab]

Can confirm.

Pi is what I use 90% of the time.

[Ash Hart @ashxhart]

@WescheNex1q @Blackwellboy :(

Tbh it seems most of the local AI boys on X are in the USA

[Ash Hart @ashxhart]

@ang3l12 @sudoingX +1 on this.

I really have to guide it along and with some fancy plugins to get it to do anything meaningful.

[Ash Hart @ashxhart]

@WescheNex1q @Blackwellboy Unfortunately not, I’m in the UK.

[Ash Hart @ashxhart]

@WescheNex1q @Blackwellboy How much? 👀

[Ash Hart @ashxhart]

@Blackwellboy Tbh I was going to sell my 256gb studio m3u untill I saw Deepseek drop.

Keeping it and trying to buy another spark.

[Ash Hart @ashxhart]

https://t.co/GR2MPbEhqS

[Poolside @poolsideai]

We've improved our serving efficiency and increased rate limits by +10x. The update is live on @OpenRouter @vercel AI Gateway and https://t.co/Pin6Re29Gr.

Thank you to everyone who has used the model, shared feedback and stuck with us while we improved the experience!

Usage is already climbing. On OpenRouter alone, Laguna S 2.1 is on pace for ~250B tokens today, 4x our daily average this week.

We are also taking 10% off our paid endpoint on OpenRouter. It's a dedicated deployment with the full 1M context window for the best performance on harder tasks.

Run Laguna S 2.1 in pool or Poolside Desktop Assistant, or plug it into @opencode, @NousResearch Hermes Agent, @kilocode, @cline or @pidotdev and let it run over the weekend.

We'll be watching the graphs.

[Ash Hart @ashxhart]

Upoladed @deepseek_ai V4 Flash 0731 in MXFP4 MLX

https://t.co/MusUgWyJn5

[Ash Hart @ashxhart]

@deepseek_ai Either benchmaxing is going on or local AI is about to cook.

[Ash Hart @ashxhart]

I've set myself a challenge to see if I can go a week without using frontier models as my go to.

Caveat: I am only using them to refine the plugins I have for @pidotdev, but overall, I am pleasantly surprised.

Here's a cheeky dashboard I put together for my Spark.

[Ash Hart @ashxhart]

@0xClandestine @jundotkim Thank you 🫡

Hopefully it opens doors for people to run larger models.

[Ash Hart @ashxhart]

PR is live for oMLX clustering; give it a try and find the bugs 🐞

May the odds ever be in your favour. 🏹

@jundotkim https://t.co/cMlekXxPRZ

[Ash Hart @ashxhart]

@0xSero What’s the spec?

[Ash Hart @ashxhart]

@lais_bsc @pidotdev Hahahahahah oh god 😂😂😂 I’m definitely renaming the plugin to DoomPoop lol.

Will do 😎

[Ash Hart @ashxhart]

@lais_bsc @pidotdev With local LLM’s they tend hallucinate or doom poop or do wild things.

I’ve created a bunch of plugins to stop the doom loop, a neural memory, chain of thought scratchpad, a proper goal system.

My best one is Cortheon which is a validation engine. I plan to OOS it soon.

[Ash Hart @ashxhart]

@pidotdev @lais_bsc Pi is my go to cli harness, and I even developed a bunch of custom plugins to help my local agents along.

Great work.