DeepSeek Open-Sources DeepInfer Inference Stack
DeepSeek's production inference engine is now open source, but the source says key internal components remain undisclosed and first-party performance claims are unaudited.
AI & ML
Frontier models, research results, agents, and developer tooling.
NVIDIA Cosmos3-H-Dreams Generates Interactive Surgical Scenes in Real Time
The model, trained on clinical cholecystectomy video and robot kinematics, is the first real-time generative surgical simulator to support live kinematic control.
Project Genie expands globally with Street View simulation
Google DeepMind is making Project Genie available to Google AI Ultra subscribers worldwide and adding Street View-powered simulation of real-world places.
AnyLanguageModel 1.0 Caps a Year of On-Device AI on Apple
The stable release consolidates a year of work on running language models locally across Apple platforms, signalling a shift toward private, offline inference.
SF Compute's Evan Conrad on Making GPU Trading Physically Deliverable
A Latent Space interview with SF Compute's CEO explains why GPU trading depends on physical delivery and an independent market operator.
DeepMind expands tools for tracing content origin and edits online
Google DeepMind is broadening its toolkit to give web users more clarity on how digital content was made and altered.
Hugging Face Backs TorchTPU for Native TPU Support
Hugging Face's support for TorchTPU gives PyTorch users a direct path to running models on Google Cloud TPUs.
Meta suspends Muse sharing after privacy leak exposes user secrets
A privacy flaw in Meta's AI assistant exposed personal and business data when users shared conversations, prompting a suspension of the feature.
AI Adoption: Tech Is First, Not the Whole Story
A new analysis argues that the tech sector's AI experience offers lessons for other industries, not a template to copy.
PaddleOCR 3.5 Brings OCR and Document Parsing to Transformers
PaddleOCR 3.5 adds a Transformers backend, making OCR and document parsing tasks accessible through the Hugging Face ecosystem.
Enterprise AI: Track cost per useful outcome, not token price
As AI moves from experiments to production infrastructure, enterprises should measure efficiency by useful business outcomes rather than raw token or GPU costs.
DeepSeek Open-Sources DeepInfer Inference Framework for V4.1-Flash
DeepSeek has released the production inference engine behind its V4.1-Flash model under an Apache 2.0 license.
DeepMind unveils Gemini for Science and agentic Gemini 3.5
Two new DeepMind releases target scientific discovery and complex workflows, signaling a push toward applied, action-oriented AI.
Nace AI's Drex Max Scores 67.59 on Computer-Use Decision Index
The new multimodal model reads text and screenshots to output calibrated probabilities for agent decisions, priced at $0.06 per million input tokens.
Samsung's Pyeongtaek mega-fab aims to reclaim AI memory lead
A visit to Samsung's expanding semiconductor complex in South Korea shows how the AI boom is reshaping the memory chip race.
Exa 3.0 Search API Claims Top Scores on 8 Self-Run Benchmarks
Exa's upgraded search API, including a cheaper Lite tier and Dynamic Highlights, reports leading results across eight benchmarks it conducted itself.
Inception Claims Mercury 3 Diffusion LLM Hits 1,068 Tokens per Second
Inception's new diffusion LLM claims a massive speed and long context, but benchmark results are self-reported.
Court orders Meta to name execs who approved AI-training torrents
A magistrate judge largely sided with authors and publishers, requiring Meta to identify decision-makers behind shadow-library downloads.