Vaibhav Srivastav

HOLY SHITT! @ZyphraAI just dropped Zonos - Apache 2.0 licensed, Multilingual, Text to Speech model with INSTANT voice cloning! 🔥

> Zero-shot TTS with Voice Cloning: Input text and a 10-30 second speaker sample to generate high-quality text-to-speech output

> Audio Prefix Inputs: Enhance speaker matching by adding an audio prefix to the text, enabling behaviors like whispering that are hard to achieve with voice cloning alone

> Multilingual Support: Supports English, Japanese, Chinese, French, and German

> Audio Quality & Emotion Control: Fine-tune speaking rate, pitch, frequency, audio quality, and emotions (e.g., happiness, anger, sadness, fear)

> Fast Performance: Runs at ~2x real-time speed on an RTX 4090

> Available on the Hugging Face Hub 🤗

Llama 3.1 8B running on Mac, 100% local, powered by llama.cpp 🔥

Two steps:

1. brew install llama.cpp

2. llama-cli --hf-repo reach-vb/Meta-Llama-3.1-8B-Instruct-Q6_K-GGUF \
--hf-file meta-llama-3.1-8b-instruct-q6_k.gguf \
-p “Sup?“ --ctx-size 8192

It's a powerful model using ~6.5GB RAM. ⚡

That's it! 🤗

Fuck yeah! Hugging Face is reproducing the FULL DeepSeek pipeline - from Data to Training pipeline FROM SCRATCH! 🔥

Join in as we create more Distilled LLMs, Data and R1-Zero⚡️

🚨 Meta released MobileLLM - 125M, 350M, 600M, and 1B model checkpoints! 🔥

Notes on the release:

Depth vs. Width: Contrary to the scaling law (Kaplan et al., 2020), depth is more critical than width for small LLMs, enhancing abstract concept capture and final performance

Embedding Sharing: Revisited and implemented embedding sharing methods ( to maximize weight utilization

Grouped Query Attention: Adopted from Ainslie et al. (2023) to optimize attention mechanisms

Immediate Block-wise Weight Sharing: Reduces latency by avoiding weight movement with minimal overhead

Performance:

> Zero-Shot Tasks: MobileLLM outperforms previous SOTA 125M/350M models by 2.7%/4.3%.

> API Calling: Comparable exact-match score to the larger LLaMA-v2 7B model 7B

Models are available on the Hub & integrated with Transformers! 🔥

Again, brilliant job from AI at Meta for continuing their Open Source + Science trend! ⚡

Yet another rewarding week in Open Source AI:

1. Google dropped Gemma 27B & 9B - The best open (commercially permissive) LLM out there, according to LYMSYS.

2. Mars5 TTS - Text to Speech with insane prosodies control & voice cloning.

3. Meta shipped LLM Compiler - beats GPT 4 on code optimisation and compiler reasoning.

4. Arcee-Spark - Qwen2 7B (w/ merging) fine-tuned further to beat GPT 3.5 on MT Bench.

5. Gemini Nano out in the wild in Chrome - On device LLM with just two lines of code (fully offline)

6. Fal released a fully Open Source GAN based Super-Resolution model (with second version already cooking)

7. NYU released Cambrian 1 - Vision Multimodal LLM that beats pretty much all other closed source competition 8-34B model size

And.. much more like Open LLM Leaderboard got a significant update, LYMSYS released Chat Vision Arena, and OpenAI released a paper on CriticGPT!

What a lovely week, can’t wait for the next to see what the community is up to! Put it down in comments if I missed something 🔥

Microsoft silently updated OmniParser on the hub 👀

60% faster than v1 - sub-second latency on a 4090! 🔥

“OmniParser is a general screen parsing tool, which interprets/converts UI screenshot to structured format, to improve existing LLM based UI agent.“

Bonus: you can try it out for free!

AlphaMaze: Teaching a 1.5B LLM to think visually and solve ARC-AGI like puzzles! 🤯

Powered by DeepSeek R1 1.5B + GRPO

All with Apache licensed checkpoints and dataset 🤗