NVIDIA's PixelUMM Merges Image and Video AI Into One 8B Model
Researchers from NVIDIA and the University of Waterloo propose a unified multimodal model that reads and writes raw pixels, skipping vision encoders and VAEs entirely.
20 stories
Researchers from NVIDIA and the University of Waterloo propose a unified multimodal model that reads and writes raw pixels, skipping vision encoders and VAEs entirely.
NVIDIA's Physis-Lang teaches video world models to reason about physics through evolving captions, lifting Cosmos 3 to the top of Physics-IQ.
Google DeepMind packaged its 740M multimodal embedding model for LiteRT, letting phones embed text, images, audio and video in one 768-dim space.
A new study tests whether populations of self-improving LLM agents actually benefit from watching each other, and finds copying crowds out exploration.
The /visualize command turns data questions into inline charts and diagrams inside the Agents Window, no separate tool required.
Vals Index v2.1 swaps in the harder Terminal-Bench 4 and adds a private Tax Agent Benchmark, reshuffling how frontier models get scored by GDP weight.
Cerebras ran Qwen3 27B at 1,500 tokens per second to book a dinner reservation in 22 seconds, 19x faster than Grok, Meta, and Claude assistants.
ARC Prize's latest chart shows collapsing per-task costs on ARC-AGI benchmarks, with reasoning models now solving problems for pennies instead of tens of dollars.
Google Research proposes a lightweight side network that steers frozen diffusion models toward better prompt alignment, beating LoRA in win rates.
Liquid AI's new d1 model returns typed decisions with calibrated probabilities in one call, generating zero output tokens for classification and routing.
OpenAI's new Decisions API turns GPT-6 Luna into a real-time classifier that picks answers from a fixed list you define.
OpenAI overhauls Codex CLI with a full-screen TUI, parallel agent management, voice input, and worktree forking for terminal-native developers.
Microsoft Research releases an open-weights 5B vision-language-action model family with embodiment-specific checkpoints and a lightweight online adaptation module.
A controller-worker split lets agents plan which partial work to extend, discard, or ship, pushing ProgramBench scores well past Codex and Claude Code.
Zyphra researchers show how sliding window layers quietly inject a recency signal into hybrid transformers, letting global attention track word order without any position encoding.
OpenAI unveiled dots at DevDay, always-on agents on GPT-6 Astra with their own cloud computer that keep projects moving between conversations.
A controlled study across 19 generation and 25 understanding tasks shows that pretraining on image generation can reduce understanding data needs by up to 10x.
Perplexity's Computer agent gains scheduled and event-triggered automations that carry memory across runs, replacing one-shot tasks with persistent background workers.
Inception Labs' new diffusion-based model returns its first spoken token in 320ms median, finally fitting a reasoning model inside the 500ms voice budget.
A 400M-parameter transformer trained only on synthetic tables predicts classification and regression in one forward pass, topping the TabArena leaderboard without any tuning.