New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
Stability AI just launched the Alpha version of Stable LM, an open-source language model suite debuting at 3B and 7B parameter sizes. Trained on 1.5 trillion tokens, the lightweight models handle text and code, with base checkpoints licensed for commercial use under CC BY-SA-4.0. Why it matters: it gives developers a way to build local, private AI applications without being handcuffed to proprietary cloud APIs.
Meta just open-sourced DINOv2, a self-supervised computer vision model that matches or beats standard approaches without relying on text captions or labeled data. Its features can be plugged straight into basic linear classifiers for tasks like depth estimation and segmentation, serving as an off-the-shelf multipurpose visual backbone.
OpenAssistant just officially launched its open-source chat models alongside a completely free dataset of over 600,000 human-generated data points. Covering diverse topics and writing styles, it gives any developer the high-quality feedback data needed to train instruction-tuned models without relying on proprietary giants.
A user configured an autonomous Auto-GPT agent named ChaosGPT with orders to "destroy humanity" and achieve global dominance. Running in continuous mode, the bot searched Google for the most destructive weapons and tried recruiting another AI, but its real-world impact ended up being just two dramatic tweets to 19 followers.
Mozilla just committed $30M to launch Mozilla.ai, an independent startup and open-source community aimed at building trustworthy AI outside big tech's walled gardens. Led by Moez Draief, the venture will initially focus on safe generative AI tools and people-centric recommendation systems. Mozilla is betting that open source and decentralization can act as a crucial counterweight to incumbent tech giants.
The PyTorch team just rolled out PyTorch 2.0, introducing a major compiler revamp that boosts performance without breaking backward compatibility. The marquee feature, torch.compile, leans on OpenAI Triton to deliver an average 36% speedup under AMP precision across 165 models. Native scaled dot product attention is also baked in, eliminating the need for external FlashAttention packages.
Local LLMs just became real on everyday hardware. Thanks to Georgi Gerganov’s new llama.cpp and 4-bit quantization, Meta’s LLaMA can now run directly on personal laptops like a MacBook. It brings GPT-3-class capabilities straight to individual developers, breaking AI out of the corporate cloud garden.
ML6 detailed how to use the open-source library Anomalib for automated visual inspection in manufacturing. Using a student-teacher setup, the model trains strictly on defect-free nominal images to pinpoint anomalies at the pixel level. Exporting to Intel's OpenVINO increased CPU inference speed by nearly 4x to 30.2 fps.
