Replicate Intelligence #9: FLUX.1, SAM 2, Gemma 2 2B, and new AI tools
Replicate's weekly bulletin covers new open-source AI releases including Black Forest Labs' FLUX.1 image generator, Meta's SAM 2 segmentation model, and Google's Gemma 2 2B language model, alongside tools for model interpretability and distributed training research.
Source: Replicate — Replicate Intelligence #9 (August 2, 2024)
Trending Models
FLUX.1
Black Forest Labs—a new company from former Stable Diffusion developers—released FLUX.1, a suite of image generation models that use an advanced “flow matching” technique. The three variants are FLUX.1 [pro] (flagship), FLUX.1 [dev] (distilled, non-commercial), and FLUX.1 [schnell] (fastest, Apache 2.0 license). The models support various aspect ratios and resolutions up to 2 megapixels and are available via Replicate.
SAM 2
Meta released SAM 2, extending its Segment Anything Model to real-time object segmentation in images and videos with zero-shot video capabilities. It is 3× faster than prior approaches. The company open-sourced the code, model weights, and the SA-V dataset (600,000 annotations on 51,000 videos). It is available on Replicate.
Gemma 2 2B
Google released Gemma 2 2B, a 2-billion-parameter open language model. It ranked above all GPT-3.5 models on the LMSYS Chatbot Arena leaderboard, though observers note it compares unfavorably on the MMLU benchmark. The weights are released under a permissive license and the model is available on Replicate.
Cool Tools
Gemma Scope
Google released Gemma Scope, an interpretability tool that uses sparse autoencoders (SAEs) to unpack information processed by Gemma 2 2B and 9B models. The release includes over 400 SAEs covering all layers.
FastHTML
Jeremy Howard, co-founder of fast.ai, released FastHTML, a Python web framework for building interactive apps in pure Python without separate HTML templates. It provides built-in support for authentication, databases, caching, and styling, and supports deployment to platforms like Railway and Vercel.
Research Radar
A new paper demonstrates federated learning for training billion-parameter language models in a distributed way without a massive centralized data center, reporting that the resulting models perform as well as centralized counterparts. The paper is available on arXiv.
The Wire · Newsletter
One careful email,
every Monday.
The week's most important AI stories, lightly edited and personally vouched for. No autoplay, no spam, easy to leave.
Comments · 0
Sign in to join the discussion.
Be the first to leave a thought.