dr.dot

Keeping the markets efficient

[dr.dot @whyYnoDr]

@DeadSeaTrader Sexy mf girlie hands, are you a hand model chick?

[dr.dot @whyYnoDr]

@Heistedxyz Redeem

[dr.dot @whyYnoDr]

Sure, there might be infinite amounts of Zec in the pool and you might never get yours out. Sure, it’s a fuckup.

But by god, this shit works SO well. It still does, ignoring all drama. I can just.. send funds, anonymously, to any chain. And it. Just. Works!

[dr.dot @whyYnoDr]

@booliontweets @kannaroy @UBERSOY1 Which was one of the first things I saw happen in a street in SF last time I visited, and still the US is pretty great. So not necessarily indicative of national success..

[dr.dot @whyYnoDr]

@Big_Tea_Rice consider selling your $PRL on https://t.co/uqHiOCfTne, it is great for spot

instant settlement, cheap fees. referral and points program running

https://t.co/DFZft3NYko

[dr.dot @whyYnoDr]

@nockusaga consider selling your $PRL on https://t.co/uqHiOCfTne, it is great for spot

instant settlement, cheap fees. referral and points program running

https://t.co/DFZft3NYko

[dr.dot @whyYnoDr]

@NeoCallss consider selling your $PRL on https://t.co/uqHiOCfTne, it is great for spot

instant settlement, cheap fees. referral and points program running

https://t.co/DFZft3NYko

[dr.dot @whyYnoDr]

@degentradingLSD It’s actually a great f-n product: https://t.co/DFZft3NYko

[dr.dot @whyYnoDr]

Best card for crypto, hands down. 3% cashback and only 1% FX spread means all foreign buys are still net positive.

UI/UX clean as well, great app @Plasma

[dr.dot @whyYnoDr]

Plasma One is by far the best off-ramp. Try it with the code: F3LREW

If you’re a high spender, DM me for more.

[dr.dot @whyYnoDr]

So Loracle, a French HYPE OG, suddenly starts dumping, and disappears from CT all of a sudden.

Again, he is French: https://t.co/j0QC3mJNK8

[0xLycurgus | Z0G arc @0xLcrgs]

he deleted his account and he keeps dumping

smells like retirement

[dr.dot @whyYnoDr]

@CSI_Starbase Definitely! Space X just did this for the 600th time as well! Happy to see Glenn starting to work as well!
https://t.co/27ZeMnfSQh

[SpaceX @SpaceX]

Falcon lands for the 600th time!

[dr.dot @whyYnoDr]

@himgajria Why do you attack me like this, out of nowhere?

[dr.dot @whyYnoDr]

@coffeecup2020 You are crushing it with these! Keep of the impressive for, very thankful for what you’re doing for us with 16GB cards

[dr.dot @whyYnoDr]

David just released an updated ”TurboQuant for weights” fitting Qwen3.5 27B on 16 GB VRAM!

Use with this custom llama for having the best of both worlds, TurboQuant for weights and TurboQuant for KV cache:

https://t.co/nhyhfkvSlK

Fits Qwen3.5 27B with large context windows on 16GB VRAM with ease 🥂

[David Y. Tan @coffeecup2020]

Turbo Quant on weight
Happy to announce TQ3_4S
2x faster, better quality than TQ3_1S, same size.

https://t.co/WXB1HsGzjh
https://t.co/lQTmtGcL41
Credit to @no_stp_on_snek who inspired me with his Turbo Quant KV. Please note Q3 wins on median PPL. Will beat this in my next model. Just need slight tweak.

[dr.dot @whyYnoDr]

@coffeecup2020 @no_stp_on_snek This is great!
I merged your llama version with the Turbo3, so we can run turboquant on kv cache simultaneously to fit a much bigger context window on 16GB VRAM:

https://t.co/3TGZZvXRd6

[dr.dot @whyYnoDr]

”TurboQuant” for both context and weights is here

TurboQuant on KV cache was already magic… but combining it with real weight size optimization was the missing piece.

@coffeecup2020 just dropped TQ3_1S weights that get near-Q4_0 quality at ~10% smaller size. Now the main problem is solved.

Here’s a clean standalone llama.cpp worktree that runs both in one single CUDA process:

- TQ3/TQ3_1S weight optimization
- Turbo3 KV cache (K + V)

Qwen3.5-27B TQ3_1S loads fully on a 16 GB card, 98k context, all layers on GPU, Turbo3 KV active. No more choosing between the two.

GitHub: https://t.co/nhyhfkvSlK

Local AI just got way more practical. Thanks to everyone pushing this forward! 🚀

[Markets & Mayhem @Mayhem4Markets]

TurboQuant is looking pretty solid. 🔥

> Original idea was to use it just for KV cache where context tokens are stored

> Now it is expanding to be used with models

> On Qwen 3.5-27B it shrinks the model down to 12.9B

> 6X memory savings vs 16-bit precision

> Stays accurate