Perplexity has open-sourced its multimodal decision model, pplx-decider-27b, alongside a new Decisions API optimized for structured, low-cost classification tasks.
This brief is compiled automatically by AI and may contain errors. Check the linked sources.
Model Releases & Updates
pplx-decider-27b & Decisions APIPPerplexity
Bottom line：Perplexity has open-sourced its multimodal decision model, pplx-decider-27b, alongside a new Decisions API optimized for structured, low-cost classification tasks.
Core Architecture：Trained to output probability distributions over fixed answer sets rather than free-form text, achieving 85.71% overall accuracy across 11 benchmarks.
Pricing：Priced at $0.04 per million input tokens with free output tokens during launch.
Bottom line：Black Forest Labs has launched FLUX 3 Image, introducing native 4K generation, bounding-box layout control, and multi-turn precise editing.
Key Capabilities：Supports targeted image modifications across multiple turns without altering surrounding pixels, alongside support for up to 10 reference images.
Availability：Commercial weights are immediately available on platforms like OpenRouter and Replicate, with open weights scheduled for release in the coming weeks.
Bottom line：Tavus has introduced Griffin, designated as the first Human Interaction Model (HIM), which reportedly passes the video Turing test.
Performance Milestone：Achieved a 48% live pass rate in the video Turing test, where conversational partners believed they were interacting with a real human.
Industry Ranking：Secured the #1 spot on NVIDIA’s benchmark for full-duplex AI video.
Bottom line：Anthropic has released Claude Code Mods, allowing developers to customize model behavior, CLI user interfaces, and subagent pipelines using simple TypeScript modules.
Integration：Mods ship inside plugins and can be easily installed via the /plugin command in the CLI or desktop application.
Built-in Safety：Includes a built-in ‘You should know’ safety mod where a background side agent monitors sessions to flag missed items.
Bottom line：Modal has officially launched VM Sandboxes, providing AI agents with fully functional Linux virtual machines complete with Docker and kernel support.
Technical Specs：Enabled simply via runtime="vm", supporting Docker, FUSE, and kernel-level features while retaining sub-second cold starts.
Scale：Early customers have already initialized over 20 million VMs for agentic workloads.
doxx.net Secures $38M Series A Led by a16zddoxx.net
Bottom line：Privacy networking startup doxx.net has raised a $38M Series A led by a16z to build Agentic Defined Networking for encrypted peer-to-peer agent communication.
Funding Details：The round includes participation from Animo Ventures and Focal.vc, valuing the infrastructure push for decentralized private networks.
Use Case：Provides secure, private communication channels for individuals and autonomous AI agents without relying on centralized intermediaries.
InstructMesh: Repairing AI-Generated 3D Models for FabricationMMIT CSAIL & Google
Bottom line：Researchers from MIT CSAIL, Google, and Northeastern University have introduced InstructMesh, combining 3D generation with LLM reasoning to let users repair and 3D print AI models using natural language.
Core Innovation：Merges Microsoft’s TRELLIS 3D generation architecture with GPT-4 reasoning capabilities to identify and correct structural flaws.
Practical Application：Enables users to issue conversational instructions to edit mesh integrity prior to physical 3D fabrication.
Bottom line：Epoch AI has partnered with YouGov to release the ChatGPT Usage Explorer, publishing usage metadata from 5,000 U.S. users spanning three years.
Dataset Scope：Covers approximately 660,000 conversations and 8.3 million messages dating back to late 2022.
Key Finding：Reveals that high-frequency usage (active 21+ days per month) quadrupled over a two-year period among active user samples.
