𝐉𝐮𝐥𝐲 18, 2024 - 𝐍𝐞𝐰 𝐀𝐈 𝐌𝐨𝐝𝐞𝐥𝐬 𝐫𝐞𝐥𝐞𝐚𝐬𝐞𝐬 A flurry of models arrived today... The idea is to make large models smaller and use synthetic data to make them as good as large models in terms of knowledge memory and acquisition. The future of LLMs is smaller and more focused, not solely larger and more AGI-like. Knowledge will become infinitely free... ⭐ 𝐃𝐞𝐞𝐩𝐒𝐞𝐞𝐤-𝐕2-0628 No.1 open-source model on the LMSYS Chatbot Arena Leaderboard. Detailed Arena Ranking: Overall No.11, Hard Prompts No.3, Coding No.3, Longer Query No.4, Math No.7. 🔗 Release: https://lnkd.in/eSUEs_hx ⭐ 𝐆𝐏𝐓-4𝐨 𝐌𝐢𝐧𝐢 The new GPT 3.5 Turbo: 60% cheaper, and outperforming it on every benchmark. It supports text and vision now, and will support "video and audio inputs and outputs" in the future. It also has a 128k context window. 🔗 Release: https://lnkd.in/evyCrqDr ⭐ 𝐌𝐢𝐬𝐭𝐫𝐚𝐥-𝐍𝐞𝐌𝐨 12𝐁 Trained in collab with NVIDIA. Outperforms Gemma2 9B and Llama3 8B. 128K context. Multilingual in 100+ languages: excels in European, Asian & Indian languages. Quant-Aware Training at FP8. Apache 2.0 License. 🔗 Release: https://lnkd.in/e_weYSv2 🔗 Release (NVIDIA): https://lnkd.in/e6b6y7Zx ⭐ 𝐒𝐧𝐨𝐰𝐟𝐥𝐚𝐤𝐞 𝐀𝐫𝐜𝐭𝐢𝐜 𝐄𝐦𝐛𝐞𝐝 𝐌 𝐯1.5 Super pragmatic text embedding for English-language search! Same small size and Apache 2.0 license as v1.0. Adds excellent vector compression (98% retrieval quality at 4% vector size vs. uncompressed). 🔗 Release: https://lnkd.in/e_7JrM9c ⭐ 𝐆𝐨𝐥𝐝𝐅𝐢𝐧𝐜𝐡 Combines the best parts of Linear Attention (via RWKV) and traditional Transformers to create something that is better than either one on its own. Hybrids are hot these days... Very capable of long context + good at recall tasks. 🔗 Abs: https://lnkd.in/ehKsdDk8 ⭐ 𝐀𝐑𝐂𝐄𝐄 𝐀𝐈 𝐌𝐨𝐝𝐞𝐥 Nova is a merge of Qwen2-72B-Instruct with a custom model tuned on a generalist dataset mixture. Approaches GPT-4 (May 2023) performance levels. Key Capabilities: Reasoning, Creative Writing, Coding, General Language Understanding. 🔗 Release: https://lnkd.in/eW5peje4 ⭐ 𝐌𝐢𝐱𝐞𝐝𝐛𝐫𝐞𝐚𝐝 𝐃𝐞𝐞𝐩𝐬𝐞𝐭-𝐦𝐱𝐛𝐚𝐢-𝐞𝐦𝐛𝐞𝐝-𝐝𝐞-𝐥𝐚𝐫𝐠𝐞-𝐯1 Open-source German/English embedding model. Fine-tuned on 30+ million pairs of high-quality German & English data. Optimized for retrieval tasks. 🔗 Release: https://lnkd.in/eANkiAeJ Pretraining scale synthetic data is the next frontier. In just 16 months, the price has dropped by 88% from GPT-3.5 Turbo-0301 to GPT-4o mini. Intelligence is becoming more affordable, making AI accessible to everyone. Cheaper tokens for everyone!
Great advice!
Thanks for sharing