Jasim

Follow Back 💯

[Jasim @Jasim85689]

ضروری نہیں کہ ہمیشہ بڑوں کی باتیں ہی درست ہوں اور سنی جائیں....

بے وقوف بھی بڑے ہو کر بوڑھے ہو جایا کرتے ہیں💯🌻

[Jasim @Jasim85689]

ضروری نہیں کہ ہمیشہ بڑوں کی باتیں ہی درست ہوں اور سنی جائیں....

بے وقوف بھی بڑے ہو کر بوڑھے ہو جایا کرتے ہیں💯🌻

[Lea 🌸🌿🌷🕊️ @tiggylea]

Good evening everyone 🧡

[Purely Green @worldclick56]

A peaceful escape into the forest. 🌲

[Travel Guide @water_diaries]

Santorini, Greece 🇬🇷

[Jasim @Jasim85689]

I Follow Back Everyone ✌️

[Miki @Mikiee_20]

If you need 10000+ followers

Say YES

I will boost your page.

[sabeer @sabeerbgfk]

ہزاروں خواہشیں ایسی کہ ہر خواہش پہ دم نکلے،
بہت نکلے مرے ارمان لیکن پھر بھی کم نکلے۔

محبت میں نہیں ہے فرق جینے اور مرنے کا،
اسی کو دیکھ کر جیتے ہیں جس پر کافر دم نکلے۔

♥️✨

[sabeer @sabeerbgfk]

تمہاری سادگی میں بھی وہ کشش ہے....

جو میری ہر نظر کو تم پر آ کر ٹھہرنے پر مجبور کر دیتی ہے۔

❤️✨

[Tahir Ai @TahirAi786]

The efficiency bar just moved again. IFM says K2-Horizon-36B-A4B is matching models that are over 20× larger, and AA ranks it #4 out of 142 comparable models, while only activating around 4B parameters per token.

[Aysel @itx__aysel]

Worth being precise here: 36B total parameters, 4B active, 25 on the AA index, and #4 out of 142 in its class. It’s not #1, but the interesting part is getting that kind of result with only 4B active parameters.

[WHALE OF AI 🐋 @WhaleOfBulls]

If you were testing a model like this, what would you look at first—the benchmark results, how the KV cache holds up with long contexts, or whether the training-stability claims actually check out?

[SYNTAX TITAN @Caromuffet]

At 1.1M steps, it’s only about 1.2–3.7 points behind the 32B dense model on MMLU, GSM8K, HellaSwag, and HumanEval. Dense still comes out ahead, but the real question is whether that small gap is worth using 8× more active compute

[Henry Cavill @Henry_cavil0]

MoE has been living in the FFN for years. MoVA basically takes that same idea and applies it to the value vectors in attention. Same concept, just a different part of the model.

[Lyr Aura @Lyr_Soul]

Cutting gradient clipping by 50% might not sound like a big deal, but it can be the difference between leaving a training run overnight and having to babysit it.

[Leo Barrett ⚡ @tysyrrr]

The FFN has been the main place we’ve been scaling sparsity for a while. Was there a specific reason attention was left alone, or was it just harder to make it work there?

[James Carter @9DDDG]

It’s a different way to think about sparsity: instead of adding more experts to the FFN, you’re putting the experts into the value computation. I’m curious to see how far this approach can actually go.

[Elora🌞 @ExpertWithElora]

No complex prompts.
No complicated workflows.

Just tell Pexo what you want — and watch the idea become a finished video. 🎬

This is vibe-creating.

[Hope Ai @Hope_Ai01]

trying to leave work and your boss hits you with “one quick task”

bro i’m already in stealth mode 😭 @flovaai

Flova made this too real