PlayCanvas creator Will Eastcott just launched SuperSplat 1.0, a free, open-source tool for inspecting and editing 3D Gaussian Splats right in your web browser. There is nothing to install—users can simply drag and drop PLY files to prune stray points and adjust scene orientation. Keep in mind it currently only supports PLY files and skips spherical harmonic rotation on export.
Google just released Gemma Scope, a massive open-source suite of over 400 sparse autoencoders to help researchers decipher Gemma 2's inner workings. By isolating more than 30 million learned features across layers, the tool functions like a microscope for mechanistic interpretability. Why it matters: Cracking open neural network black boxes is key to catching hallucinations, deception, and safety vulnerabilities before they cause harm.
Activision just released an open-source dataset pulled from Call of Duty: Warzone’s Caldera map on GitHub. Formatted in OpenUSD, it includes near-complete map geometry and anonymized player movement traces to support research. It marks one of the largest production-validated gaming datasets ever made public, giving generative AI models rich 3D environments to learn from.
The White House just pushed back against clamping down on open-source AI. A new Commerce Department report found no need to restrict developers from releasing model weights, citing insufficient evidence of major harms and real benefits to open innovation. But officials warn they're still monitoring safety and will step in if bigger risks appear.
Meta just dropped its open-source Llama 3.1 family, headlined by a massive 405B model that rivals GPT-4o and Claude 3.5 Sonnet across key benchmarks. The entire lineup supports a 128K context window, and Meta has updated its license so developers can finally use 405B outputs to distill smaller models.
Apple's research team just released open DCLM models with 7B and 1.4B parameters on Hugging Face, sharing weights, code, and training datasets. The 7B model scored 63.7% on MMLU, outperforming Mistral-7B while using 40% less compute. It shows Apple's data curation prowess, though it's research-focused rather than on-device software.
Mistral AI just launched Codestral Mamba, a code-focused model built with Albert Gu and Tri Dao that swaps Transformers for the Mamba architecture. It delivers linear-time inference and handles tested in-context retrieval up to 256k tokens without slowdowns. Even better, Mistral open-sourced it under Apache 2.0 for free local deployment.
Stability AI just revamped its Community License following heavy community pushback over Stable Diffusion 3. Under the revised terms, individual creators and small businesses making less than $1M in annual revenue can now use SD3 Medium and other models commercially for free, with zero output caps. The team also admitted SD3 Medium fell short on quality and promised an update soon.
French research lab Kyutai just unveiled Moshi, an open-source conversational voice AI boasting an end-to-end latency of only 200 milliseconds. Built from scratch in six months, it can listen and talk at the same time and run directly on a laptop without an internet connection.
Kwai's research team just open-sourced LivePortrait, an efficient portrait animation model that brings still human or animal photos to life using driving videos. Unlike diffusion models, it uses implicit keypoints to clock in at just 12.8ms per frame on an RTX 4090.
Shanghai AI Lab just open-sourced InternLM 2.5-7B, cramming a massive 1M token context window into a lightweight architecture. The model claims top scores for sub-12B models on Hugging Face's Open LLM Leaderboard, bringing heavy-duty reasoning and tool use to local hardware.
Google just officially released Gemma 2, introducing ultra-efficient 9B and 27B open models for developers. The 27B variant punches above its weight by matching models twice its size while running full precision on a single Nvidia H100, and the 9B edges out Llama 3 8B. It is available now under a commercial-friendly license.
Microsoft just released Florence-2 on Hugging Face under a permissive MIT license, cramming multi-task vision capabilities into 232M and 771M parameter models. It uses text prompts to handle object detection, captioning, and segmentation, outperforming models a hundred times its size in zero-shot tests. This gives developers a single, lightweight tool instead of juggling multiple specialized vision pipelines.
Meta's FAIR team just released five new AI research models and tools to the public. Highlights include Chameleon for mixed text-and-image generation, JASCO for chord- and beat-controlled music synthesis, and AudioSeal, an audio watermarking system that detects synthetic speech up to 485 times faster.
Apple just released 20 new Core ML models and four datasets on Hugging Face, including FastViT and DepthAnything. Built specifically to run on-device without an internet connection, they let developers build fast computer-vision features while keeping user data strictly private.
DeepSeek just open-sourced DeepSeek Coder V2, an MoE-based code model supporting 338 languages and a 128K context window. It beat GPT-4 Turbo, Claude 3 Opus, and Gemini 1.5 Pro on major coding benchmarks like HumanEval and MBPP+, marking a rare open-source victory over leading proprietary models. Available under the MIT license in 16B and 236B sizes.
The Qwen team just launched Qwen2, an open-source model family spanning five sizes from 0.5B to 72B parameters. Trained on 27 additional languages, the flagship Qwen2-72B goes toe-to-toe with Llama-3-70B across major benchmarks, while context lengths stretch up to 128K tokens. Except for the 72B tier, all smaller models are now released under the Apache 2.0 license.
