@nenad1001: #jelenakarleusa #karleusa #jklive #foryou #fyp

Nenad
Nenad
Open In TikTok:
Region: RS
Saturday 29 August 2026 21:28:17 GMT
5968
290
3
31

Music

Download

Comments

emiii9804
Emiii :
2026-08-29 21:35:14
0
user2998562153060
user2998562153060 :
🥰🥰🥰
2026-08-29 21:31:55
0
To see more videos from user @nenad1001, please go to the Tikwm homepage.

Other Videos

This AI model was given a research paper and was tasked to reproduce the entire thing itself. 12 hours later, it created 18 commits, 23 figures, and replicated every core experiment. The model isn't GPT, Claude or Gemini. It's Minimax M3, an open-weight frontier model featuring a 1 million token context window and native multimodality. It processes queries efficiently using Minimax sparse attention (MSA), which is an extension of grouped-query attention (GQA) where multiple query heads share one key-value head. MSA adds 2 small projection matrices that form an index branch to choose which past tokens each query should compute scores against. Keys and values are split into blocks of 128 tokens. The index branch scores every block by its highest-scoring token and keeps the top 16 blocks. Attention is then run over those chosen blocks. Compared to dense attention, MSA computes only on a subset of the full tokens. At 1 million tokens, that's 14x faster prefill and 28x less attention FLOPs with comparable quality to full attention. Minimax M3 can be used in coding harnesses like Claude code, Codex, OpenCode, and more. You only need an API key, which can also use token plans. M3 can reason across long, multimodal contexts at once while generating and modifying the code. It excels in multimodal applications and follows UI systems really well at a low token cost. If M3’s native multimodality and long context sounds interesting, check the link in my bio to try MiniMax M3. #MiniMax #MiniMaxM3 #MiniMaxCode
This AI model was given a research paper and was tasked to reproduce the entire thing itself. 12 hours later, it created 18 commits, 23 figures, and replicated every core experiment. The model isn't GPT, Claude or Gemini. It's Minimax M3, an open-weight frontier model featuring a 1 million token context window and native multimodality. It processes queries efficiently using Minimax sparse attention (MSA), which is an extension of grouped-query attention (GQA) where multiple query heads share one key-value head. MSA adds 2 small projection matrices that form an index branch to choose which past tokens each query should compute scores against. Keys and values are split into blocks of 128 tokens. The index branch scores every block by its highest-scoring token and keeps the top 16 blocks. Attention is then run over those chosen blocks. Compared to dense attention, MSA computes only on a subset of the full tokens. At 1 million tokens, that's 14x faster prefill and 28x less attention FLOPs with comparable quality to full attention. Minimax M3 can be used in coding harnesses like Claude code, Codex, OpenCode, and more. You only need an API key, which can also use token plans. M3 can reason across long, multimodal contexts at once while generating and modifying the code. It excels in multimodal applications and follows UI systems really well at a low token cost. If M3’s native multimodality and long context sounds interesting, check the link in my bio to try MiniMax M3. #MiniMax #MiniMaxM3 #MiniMaxCode

About