🚀Introducing OpenChat 3.6 🌟Surpassed official Llama3-Instruct—with 1-2M synthetic data compared to ~10M human labels 🤫GPTs are close to limits—excel at generation but fall short at complex tasks 🎯We are training next gen—capable of deterministic reasoning and planning 🔗

Will Sudoku become the MNIST for reasoning? Simple rules, clear structure, unique solutions—yet surprisingly challenging for modern LLMs, often requiring explicit trial-and-error to solve. https://t.co/BmqUaW8Qno

🚀Introducing Hierarchical Reasoning Model🧠🤖 Inspired by brain's hierarchical processing, HRM delivers unprecedented reasoning power on complex tasks like ARC-AGI and expert-level Sudoku using just 1k examples, no pretraining or CoT! Unlock next AI breakthrough with

Thrilled to see RSP featured at AAAI'25! This pioneering concept was a key inspiration for developing OpenChat! 🚀 #AI #AAAI25

skronge bones in that one 🔍 excellent job! a 7b model out cracking gpt4 turbo and gpt4o and claude 3 sonnet!

4)This is why I am embarking on a journey to explore new frontiers in AI, specifically targeting the current limitations of GPTs in Planning and Reasoning.

3) However, while training these new models, I can't help but realize the upper limit of what autoregressive models can do. They struggle to solve complex tasks such as software engineering, advanced mathematics, and creating super assistants. It is mathematically challenging for

2) In Openchat 3.6, we pushed Llama3 8B to a new level of performance while retaining the flexibility for further SFT, so developers can better tailor our model for each unique use-cases

1）We developed a new continous pre-training method, Meta-Alignment, for LLMs which achieves similar results to extensive RLHF training that Meta did with Llama3 Instruct. This process is both data and compute efficient using primarily synthetic data at less than 10% of data set

🚀Introducing OpenChat 3.6 🌟Surpassed official Llama3-Instruct—with 1-2M synthetic data compared to ~10M human labels 🤫GPTs are close to limits—excel at generation but fall short at complex tasks 🎯We are training next gen—capable of deterministic reasoning and planning 🔗

🚀 The World's First Gemma fine-tune based on openchat-3.5-0106 data and method (C-RLFT). Almost the same performance as the Mistral-based version. 6T tokens = secret recipe? HuggingFace:

🚀Kudos to @huggingface ! OpenChat-3.5 Update 0106 has landed on HuggingChat & Spaces! Explore now! Experience open-source AI at ChatGPT & Grok level! 🤗 HuggingChat: https://t.co/ic5jOEYlUm 🌌 Spaces: https://t.co/ov8AdxWSgD 🖥️ OpenChat UI:

🚀Announcing OpenChat-3.5 Update 0106: 𝗪𝗼𝗿𝗹𝗱’𝘀 𝗕𝗲𝘀𝘁 𝗢𝗽𝗲𝗻 𝗦𝗼𝘂𝗿𝗰𝗲 𝟳𝗕 𝗟𝗟𝗠! Experience ChatGPT & Grok-level AI locally 💿! Surpassing Grok-0 (33B) across all 4 benchmarks and Grok-1 (???B) on average and 3/4 benchmarks 🔥. 🎯 This update mainly enhanced

(3/3) Full comparison with Grok

(2/3) Coding scores on HumanEval+

(1/3) Additional benchmark results, including ChatGPT and other open-source models

🚀Announcing OpenChat-3.5 Update 0106: 𝗪𝗼𝗿𝗹𝗱’𝘀 𝗕𝗲𝘀𝘁 𝗢𝗽𝗲𝗻 𝗦𝗼𝘂𝗿𝗰𝗲 𝟳𝗕 𝗟𝗟𝗠! Experience ChatGPT & Grok-level AI locally 💿! Surpassing Grok-0 (33B) across all 4 benchmarks and Grok-1 (???B) on average and 3/4 benchmarks 🔥. 🎯 This update mainly enhanced

We achieved almost ideal MoE fine-tuning performance (equivalent to a dense model with the same active parameters). <10% overhead