MIT's MOSAIC Beats the Best Chemistry AI by 23 Points
A single flow matching transformer learns molecules and three spectra jointly, beating specialist spectra-to-structure baselines by 23.2 points on QM9S.
20 stories
A single flow matching transformer learns molecules and three spectra jointly, beating specialist spectra-to-structure baselines by 23.2 points on QM9S.
Tufa Labs' tufazip squeezes a billion bytes of Wikipedia into roughly 105 MB using a Transformer as its only predictor, taking #2 on a 20-year-old benchmark.
A 4.5M-parameter English TTS model runs 70x real time on CPU, trained from scratch in under 8 hours on a single RTX 4090
Cloudflare open-weights a redesigned 27B Clairic and new 9B Clairic-flash under Apache 2.0, adding PDF support and a parallel slot generation architecture tuned for typed data extraction.
A community ComfyUI workflow packages MiniMax H3 video generation with a Turbo LoRA and resolution controls, letting modest local GPUs run the open-weight model.
Apple's new depth model handles single images, video, and live feeds at native resolution, fixing the sharpness-versus-jitter tradeoff plaguing monocular depth.
RadixArk's first open-weight model takes Qwen3.8-27B and more than doubles its Terminal-Bench 4.1 score using pure RL on 1,200 agent tasks.
Freepik released an open source NSFW image classifier on Hugging Face with four severity levels, pip packaging, and benchmarks against existing detectors.
LLM cascades let a small model like DeepSeek V4.1 Flash handle work alone, escalating only 22% of cases to Claude Opus 5.5 while cutting cost per 1,000 cases from $3.86 to $0.88
Meta and Oxford's 1B-parameter transformer reconstructs full 3D scenes from one to hundreds of images in under a second, no optimization required.
Black Forest Labs' distilled 9B image model lands in ComfyUI as a single safetensors file, generating and editing images in under a second on consumer GPUs.
OpenCode Console now accepts short-lived OIDC tokens from CI runners, letting workflows call models without storing long-lived API keys as secrets.
A community INT4 quantization of Gemma 2 9B Instruct crosses 800K downloads, shrinking the model to around 6 GiB of VRAM for high-throughput serving.
A Rust and Tauri desktop app wraps the YuE2 song model with an editable score, covers, LoRA training, and a local MCP server.
The /btw slash command grows beyond quick side questions, letting you fork a conversation, explore tangents, and promote the ones worth keeping.
Chan Zuckerberg Biohub is now hosting the ESM-2 650M protein language model on Hugging Face, continuing stewardship of the ESM lineage after Meta's team spun out.
Vals AI opened a public registry of model cheating, reward hacking, and rule-bypass episodes, seeded with 15 incidents from its own evaluations and open to community submissions.
Perplexity's agent now runs inside any Chromium browser through the Chrome extension, taking over tabs from a side panel without the Comet install.
Google's agentic IDE Antigravity now ships an official MongoDB Atlas plugin with seven skills covering natural language queries, schema review, and Vector Search.
vLLM ships day-0 kernels for Thinking Machines' Inkling 2 Mini, a 26B looped transformer that runs its middle 12 layers three times per token.