Open-Source AI Surges: K2-Think 32B, Unsloth’s <15GB RL, and GGUF MoE Models Released
Model Releases & Updates
K2-Think 32B - UAE’s Open-Source Reasoning Model Released
The UAE has released K2-Think 32B, an open-source reasoning model with strong benchmark performance and high token generation speed. The model is available on Hugging Face alongside a detailed research paper.
InclusionAI’s 103B MoE Models Now Available in GGUF Format
InclusionAI has released GGUF versions of its Ring-Flash 2.0 (Reasoning) and Ling-Flash 2.0 (Instruct) Mixture-of-Experts (MoE) models, improving accessibility for local deployment.
AI Training & Compute Trends
Epoch AI Research: GPT-5 Used Less Compute Than GPT-4.5, GPT-6 Expected to Reverse Trend
Epoch AI Research reports that GPT-5 required less overall compute for training compared to GPT-4.5, but predicts GPT-6 will increase compute usage again, signaling a potential shift in AI training resource allocation.
AI Optimization & Local Deployment
GLM-4.5 355B Now Runs on RTX 3090 GPUs with 128K Context
Users can now run the 355B-parameter GLM-4.5 model with 128K context on RTX 3090 GPUs, showcasing advancements in making large models more hardware-efficient and accessible.
Unsloth Enables Faster Reinforcement Learning with GPT-OSS (<15GB VRAM)
Unsloth has optimized Reinforcement Learning (RL) for GPT-OSS, achieving faster inference speeds and <15GB VRAM usage, improving efficiency for training and deploying AI models locally.
AI Product & Feature Updates
Perplexity AI September Updates: Email Assistant, Flashcards, Stock Indicators & More
Perplexity AI rolled out new features, including an Email Assistant for Max subscribers, Language Learning Flashcards, Stock Indicators on iOS, and Image Model Selection on iOS, enhancing productivity and user experience.