
AssemblyAI's Universal-3.5 Pro Gives Voice Agents Memory Between Turns
AssemblyAI's Universal-3.5 Pro Realtime brings conversation memory to streaming STT, letting voice agents feed prior turns to the model mid-call without reconnecting.
20 stories

AssemblyAI's Universal-3.5 Pro Realtime brings conversation memory to streaming STT, letting voice agents feed prior turns to the model mid-call without reconnecting.

Apptronik's 90,000 sq ft Robot Park and Apollo 2 form a live data pipeline feeding Google DeepMind's Gemini Robotics foundation model, with Apollo 3 in the crosshairs.

Tencent's Hy3 is a 295B MoE open-weight model that rivals trillion-parameter flagships at a fraction of the compute cost, now free on OpenRouter for two weeks.

Bespoke Labs raises $40M to build the RL training environments that could finally make AI agents reliable enough for real enterprise work

Sarvam's new MCP server lets coding agents discover and use its Indian-language AI models with a single config line

A 5B parameter latent diffusion model simulates a full 2v2 Rocket League match in real time at 20 FPS from player key presses.

Runway opens a Paris research hub with a $30M initial investment, doubling down on world models and physical AI in Europe's hottest AI city.

Unsloth ships lossless GGUF builds of DeepSeek-V4-Flash, making the 284B MoE model locally runnable on 110GB RAM after a critical llama.cpp fix landed July 7th.

daily-arXiv-ai-enhanced auto-crawls arXiv papers daily, summarizes them with DeepSeek, and serves a personalized reading interface — all for free using GitHub Actions and Pages

Sakana AI launches a free Japanese translation tool powered by Namazu, its Japan-adapted model, with translate, proofread, and Q&A modes

F5-TTS clones any voice from 5 seconds of audio using flow matching, no phoneme alignment needed, fully open-source and MIT licensed

GitHub Copilot CLI now supports custom agents defined in Markdown, turning one-off terminal prompts into repeatable, team-shareable workflows for security, releases, and more.

Tencent open-sources CubeSandbox, a Rust-built microVM sandbox for AI agents that cold-starts in under 60ms with hardware-level isolation and drop-in E2B compatibility

Mission Control is a free, self-hosted dashboard for running AI agent fleets with zero external dependencies, built on SQLite and Next.js

Tencent's 295B MoE open-weights model rivals trillion-parameter flagships on agentic tasks at a fraction of the inference cost

Mistral's Leanstral 1.5 solves 587/672 Putnam problems at ~$4 each, finds 5 unreported open-source bugs, and scales smoothly to 4M tokens

DeepSeek's open-source 7B chat model, trained on 2 trillion tokens, brings bilingual conversational AI to any GPU with a commercial-friendly license.

DeepSeek's 7B base model, trained on 2T tokens from scratch, is a commercially free, bilingual foundation for fine-tuning and research

DeepSeek's 67B base model lands on Hugging Face with full weights, commercial license, and plug-and-play support for vLLM, SGLang, and Transformers

DeepSeek releases DeepSeekMoE 16B, a sparse model matching LLaMA2 7B quality at 40% of the compute cost