🌸¥a🌸

🌸 Just a girl with a happy heart

[🌸¥a🌸 @do0ubleP]

꧁༺ #بدايـــــــــــــــہ ༻꧂

@puenerzuile
♡ 750♡

@rafaelioramos
♡ 700 ♡

@marrrli
♡ 830♡

@spameriaa
♡ 700♡

@do0ubleP
♡ 660♡

@pconelg
♡ 800♡

بالتوفيق للجميع 🥀💫

꧁༺ 𝕏 ༻꧂

[Aakash Kanojiya @Kanojiyaaakash1]

Want to build your own coding-agent UI?

Ante's mini-tui example is one Rust file: streaming replies, tool activity, y/n approvals, and Esc to interrupt.

The agent runs underneath.

Start there and make the interface yours

[NASMA ELLIS @nasmaemad94]

What stands out to me is how much they’re squeezing out of just ~4B active parameters.

You’re getting performance surprisingly close to a 32B dense model, without blowing up the KV cache, while also making training more stable.

It feels less like just another MoE tweak and more like a different direction for scaling sparsity.

[SHARI @shari1345]

If you’ve ever had to size a serving cluster around KV-cache usage, you’ll know why the “no extra KV-cache cost compared with GQA” part is probably the most important detail here.

[Mr bean @Bean_i_9]

Mixture-of-Value Attention basically moves the sparsity into the value part of multi-head attention, rather than putting it in the feed-forward block.

[ZARA NISA @Zaranisa88]

Each token only uses 4 out of the 64 value experts, so about 6.25% of the pool is active. The value vectors keep the same dimensions, which means the KV cache stays about the same size as standard GQA.

[SCARLETT AI✨ @Beast_85]

The pre-training loss curve is pretty close to their 32B dense model, even with only ~4B active parameters per token. The dense model is still ahead, but the gap isn’t that big

[SAIRA 🎀 @_Itx_Saira_]

A lot of architecture papers propose ideas that don’t play nicely with FlashAttention. This one is different—it keeps FlashAttention, GQA, and sparse attention all in play.

[🌸¥a🌸 @do0ubleP]

@n8333n @100seatsxyz Will watch closely for updates on 27th

[Robinhood Girl @n8333n]

no telegram. no alpha group. no private channel before launch.
everything about $SEAT is posted in one place: @100seatsxyz.
if you're not following it, you're not early. you're just not informed.

[Zohaib Ai @AiwithZohaib]

K2-Horizon-36B-A4B shows how AI scaling is evolving beyond simply increasing parameter counts.

With 36B total parameters but only ~4B active per token, it aims to deliver strong reasoning while keeping inference more compute-efficient.

The standout piece is MoVA — introducing sparsity into the value path of attention, alongside the traditional MoE architecture in the FFN.

A fascinating direction for building larger, more capable AI systems without scaling compute requirements linearly.

[Mr Alexander Official @wiigiii_]

The efficiency game is moving fast.

K2-Horizon-36B-A4B is reportedly delivering performance comparable to models 20× larger, according to IFM.

Artificial Analysis places it #4 among 142 comparable models, while only ~4B parameters are activated per token.

A strong example of how smarter architecture can push AI performance without scaling compute linearly.

[ZENZOY @zordeitwz]

MoE introduced sparsity into the feed-forward layers, and MoVA takes that concept further by bringing expert routing into the value computation of attention.

What’s interesting here is the new dimension of scaling: rather than simply adding more FFN experts, @IFM_AI is exploring how model capacity can expand through the attention mechanism itself.

A promising architectural direction for building more capable models while keeping compute efficient.

[NoorAI @noorwithwifi]

Putting @MiniMax_AI H3 through its paces - seamless omni reference, dynamic motion graphics, and crisp typography control all in one workflow. 🎬

Think you can cook up something insane? Enter the MiniMax H3 x Picsart Challenge!

• $50,000 in prizes 🎁
• Submissions end Sep 23 ⏳
• Join now: https://t.co/TehYAhtxKm

cc: @Picsart

[NoorAI @noorwithwifi]

“Can we see an example?”

That was the example.

[Saeed Anwar AI @PakisTa12503890]

Claude can now build your entire mobile app, like an Apple-level $350K dev, in minutes and FOR FREE.

What used to require a full team and weeks of work can now be done with just a few well-thought.

Want the full guide??👇

Comment "Need" . I'll send you Everything

[NASMA ELLIS @nasmaemad94]

AI generation doesn't have to be expensive to experiment with.

@Flovaai is offering up to 80% off + 365-day locked rates.

Seedance 2.5 from $0.022/sec

[ZARA NISA @Zaranisa88]

Wan 3.0 Prime from $0.022/sec.

Image 2.5 from $0.011/image.

Seedance 2.5 from $0.022/sec.

@Flovaai is making AI generation much more affordable.

[Leo Barrett ⚡ @tysyrrr]

AI creators: this pricing is worth checking 👀

@Flovaai is offering up to 80% off.

Seedance 2.5 480p from $0.022/sec.

[🌸¥a🌸 @do0ubleP]

@HalimaAI_ Great

[🌸¥a🌸 @do0ubleP]

@feelarggg Great