Qwen released QvQ 72B OpenAI o1 like reasoning model on Hugging Face with Vision capabilities - beating GPT4o, Claude Sonnet 3.5 🔥

BOOOOM! Meta released Llama 3.3 70B - 128K context, multilingual, enhanced tool calling, outperforms Llama 3.1 70B and comparable to Llama 405B 🔥 Comparable performance to 405B with 6x LESSER parameters ⚡

Introducing Indic-Parler TTS - Trained on 10K hours of data, 938M params, supports 20 Indic languages, emotional synthesis, apache 2.0 licensed! 🔥 w/ fully customisable speech and voice personas! Try it out directly below or use the model weights as you want! 🇮🇳/acc

you can just do things - ask AI to create your SQL queries and execute them right in your browser! 🔥 let your creativity guide you - powered by qwen 2.5 coder 32b ⚡ available on all 254,746 public datasets on the hub! go check it out today! 🤗

This demo of structured data extraction running on an LLM that executes entirely in the browser (Chrome only for the moment since it uses WebGPU) is amazing My notes here: simonwillison.net/2024/Nov/29/...

Fuck it! Structured Generation w/ SmolLM2 running in browser & WebGPU 🔥 Powered by MLC Web-LLM & XGrammar ⚡ Define a JSON schema, Input free text, get structured data right in your browser - profit!!

FYI, here's the entire code to create a dataset of every single bsky message in real time: ``` from atproto import * def f(m): print(m.header, parse_subscribe_repos_message()) FirehoseSubscribeReposClient().start(f) ```

I have converted a portion of my NLP Online Masters course to blog form. This is the progression I present that takes one from recurrent neural network to seq2seq with attention to transformer. mark-riedl.medium.com/transformers...

I'm disheartened by how toxic and violent some responses were here. There was a mistake, a quick follow up to mitigate and an apology. I worked with Daniel for years and is one of the persons most preoccupied with ethical implications of AI. Some replies are Reddit-toxic level. We need empathy.

yo! nvidia finally released the weights for Hymba-1.5B - outperforms Qwen, and SmolLM2 w/ 6-12x less training trained ONLY on 1.5T tokens > massive reductions in KV cache size and improved throughput > combines Mamba and Attention in a hybrid parallel architecture with a 5:1 ratio and meta-tokens

Let's go! We are releasing SmolVLM, a smol 2B VLM built for on-device inference that outperforms all models at similar GPU RAM usage and tokens throughputs. SmolVLM can be fine-tuned on a Google collab and be run on a laptop! Or process millions of documents with a consumer GPU!

Smol TTS keeps getting better! Introducing OuteTTS v0.2 - 500M parameters, multilingual with voice cloning! 🔥 > Multilingual - English, Chinese, Korean & Japanese > Cross platform inference w/ llama.cpp > Trained on 5 Billion audio tokens > Qwen 2.5 0.5B LLM backbone > Trained via HF GPU grants

SmolLM - run, pre-train, fine-tune, evaluate SoTA fully open source LM 🔥 Run with Transformers, MLX, Transformers.js, MLC Web-LLM, Ollama, Candle and more! Apache 2.0 licensed codebase - go explore now!

Massive week for Open Source AI/ ML Mistral Pixtral & Instruct Large - ~123B, 128K context, multilingual, json + function calling & open weights Allen AI Tülu 70B & 8B - competive with claude 3.5 haiku, beats all major open models like llama 3.1 70B, qwen 2.5 and nemotron

Apple released blazingly fast CoreML models AND an iOS app to run them on iPhone! ⚡ > S0 matches OpenAI's ViT-B/16 in zero-shot performance but is 4.8x faster and 2.8x smaller > S2 outperforms SigLIP's ViT-B/16 in zero-shot accuracy, being 2.3x faster, 2.1x smaller, and trained with 3x fewer data

Check out my new swanky handle! 🦋 - Drop your Hugging Face ID in the comments if you want the same!

LFG!! XGrammar: a lightning fast, flexible, and portable engine for structured generation! 🔥 > Accurate JSON/grammar generation > 3-10x speedup in latency > 14x faster JSON-schema generation and up to 80x CFG-guided generation GG MLC team is literally the best in the game and slept on! ⚡

🚨 UPDATE: New Whisper based model competing with Nvidia on Open ASR Leaderboard! 🔥 CrisperWhisper aims to transcribe every spoken word exactly as it is, including fillers, pauses, stutters and false starts Whisper Large V3 fine-tune - beats it by roughly ~1 WER margin ⚡ hf.co/spaces/hf-au...

OH WOW! The Whale aka DeepSeek is BACK!! New model, with complete reasoning outputs and a gracious FREE TIER too! 🔥 Here's a quick snippet of it searching the web for the right documentation, creating the JS files plus the necessary HTML all whilst handling Auth too ⚡

Great day for M/LLMs, just released Mistral & Pixtral Large - ~123B, 128K context, Multilingual, JSON + Function calling support & open weights! 🔥 Pixtral Large: huggingface.co/mistralai/Pi... Mistral Large: huggingface.co/mistralai/Mi...

New spaces of the week! 🔥 > Qwen 2.5 Coder Artifacts > Flux Kolors Character > X Potrait > Text Behind Image 🤯 > DimensionX > MagicQuill > JanusFlow 1.3B > Netflix Recommentation Check them out at hf.co/spaces 🏃

What a brilliant week in Open Source!

🚨 Nexusflow released Athene v2 72B - competitive with GPT4o & Llama 3.1 405B Chat, Code and Math 🔥 > Arena Hard: GPT4o (84.9) vs Athene v2 (77.9) vs L3.1 405B (69.3) > Bigcode-Bench Hard: 30.8 vs 31.4 vs 26.4 > MATH: 76.6 vs 83 vs 73.8 Open science ftw! ⚡

Smol TTS models are here! OuteTTS-0.1-350M - Zero shot voice cloning, built on LLaMa architecture, CC-BY license! 🔥 > Pure language modeling approach to TTS > Zero-shot voice cloning > LLaMa architecture w/ Audio tokens (WavTokenizer) > BONUS: Works on-device w/ llama.cpp ⚡

Our multimodal team is releasing IDEFICS! One year in the making, the first open SOTA visual LLM: images/text input, text output. Think multimodal ChatGPT! We release 2 sizes: 80B🐳 & 9B🐿️ Read: https://huggingface.co/blog/idefics Try: https://huggingface.co/spaces/HuggingFaceM4/idefics_playground

What a fantastic and 🐌 paced weekend! ♥️

Kinda hilarious that my Twitter feed is burning with HF and GitHub being down. At Bsky, people are sharing memes and cat pictures! ✨