Meta

Latest AI news, models and releases from Meta. ['Astryx', 'Bonsai-27B', 'Business Agent', 'BYT5', 'CodeLlama', 'expressive voice AI', 'FAIRChem v2 UMA', 'Gemma 4', 'Gemma4:e2b', 'HuBERT', 'llama', 'Llama', 'LLaMA', 'Llama-2 70B', 'LLaMA2-chat-70B', 'Llama 3', 'Llama 3.1', 'Llama 3.1-405B', 'Llama-3.1-405B', 'LLaMA-3.1 405B', 'Llama 3.1-70B', 'Llama-3.1-70B', 'Llama 3.1 8B', 'Llama 3.1-8B', 'Llama-3.1-8B', 'Llama 3.2', 'LLaMA 3.2', 'Llama-3.2-1B-Instruct', 'Llama 3.3', 'llama-3.3-70b', 'Llama 3.3-70B', 'Llama 3 8B', 'Llama3-8B-Instruct', 'Llama 4', 'Llama 4 Maverick', 'LLaMA 65B', 'Llama 70B', 'LLaVA', 'M2M100', 'M**a AI']

Official site ↗

Research 🇷🇺

Token Cramming: Ideal Text Compression Does Not Preserve Meaning – AIRI Study

Researchers at AIRI demonstrated that the token cramming method, which compresses up to 1568 tokens into a single embedding, does not guarantee meaning preservation: under greedy decoding, accuracy drops from 99.96% to 40%. A new method, progressive cramming, was proposed to achieve 100% reconstruction, but it was found that even a perfectly compressed embedding impairs the model's reasoning ability, and the cause lies in the early layers of the transformer.

Every Eval Ever Results Now on Hugging Face Model Pages

The EEE (Every Eval Ever) project, launched by the EvalEval Coalition, standardizes AI evaluation reporting using a unified JSON schema. Now integrated with Hugging Face Community Evals, it enables cross-linking between model pages and full evaluation records, with over 229,000 results from 31 formats already collected.

Hugging Face