😍Awesome work by JinaAI! They just dropped Jina CLIP v2, a 0.9B multilingual multimodal embedding model. The model combines Jina XLM-RoBERTa (561M params) for text and EVA02-L14 (304M params) for vision.

Official announcement: jina.ai/news/jina-cl...

#Embedding #RAG #multilingual #multimodal

Exciting breakthrough by the Salesforce team! 🚀 They've developed Moirai-MoE, the first time-series foundation model.

What's mindblowing? It beats GPT-4 in time-series tasks despite being 1000x smaller! 🤯 That's what I call innovation.

#LLM #MOE #Timeseries

🔥Incredible work by the thu-ml team on making attention mechanisms faster and lighter!
They've created a 4/8-bit quantization approach that's a drop-in replacement for Flash Attention.

⭐️Key results:
- 3x faster than Flash Attention 2
- Maintains 99% of original performance

#LLM #Efficient

🥳Awesome work by @nyra_health! They made CrisperWhisper, a significant upgrade to Whisper's speech transcription capabilities. This could be a game-changer for opensource speech processing.

⚡️It's #2 on OpenASR Leaderboard

Github: github.com/nyrahealth/C...

#AI #OpenSource #Speech

🎯 Just tried LearnLM on Google's AIStudio! This Gemini-based model is specifically tuned for education.

To be honest, this is the best model can give structured output I've ever tried.

Key features:
- Multi-modal learning support
- Wonderful voice tone
- Perfect structured ourput

#EdLLM #AI

We have a new model on Livebench from StepFun. This is a trillion-parameter Mixture of Experts (MoE) language model.

Yes you read that right. It's a TRILLION-PARAMETER 🤯

#LLM #MOE #AI #AGI

Official Annoucement: stepfun.com#step2
Livebench: livebench.ai#/