Deepgram acquires Daily, the company behind the open source voice AI framework Pipecat, for an undisclosed sum. The companies say Pipecat will remain open source. Deepgram and the Pipecat team also release the Real-Time Context Standard, an open protocol for voice agents.
Dream, a startup founded by former Palantir executives Andy Elder and Arjun Mathur, exits stealth with $17 million in seed funding led by Andreessen Horowitz. Its built world model, trained on more than 100 million square feet of permit data, designs structurally precise buildings. The company says the platform is deployed on 10 projects in Texas and California.
Mistral releases the weights for Mistral Large 4, a trillion-parameter open-weight model available on Hugging Face and through its API, letting businesses and governments run it on their own systems for cyber defense. Mistral says the model uses 47 percent fewer tokens per task than GLM-5.3 and outranks other open models on Artificial Analysis' Coding Agent Index.
GitHub says Copilot Automations can now learn from their outcomes and propose changes to their own prompts. Humans review each proposed change before it takes effect. The feature requires Copilot Pro, Pro+, Max, Business, or Enterprise.
Google launched a preview of its Connected Apps Platform, a single console where developers build once for both Search and the Gemini app. Developers integrate through a cloud-hosted MCP server or Android MCP and register in the Google Connected Console. Google said partner apps from Adobe, Experian, Localiza and Peloton are live in Gemini.
Google Cloud launches Agent Compute Engine, a microVM service giving each Gemini agent a hardware-isolated sandbox, with preview in the coming months. The company says idle agents pause through checkpoint and restore. Its open-source Agent Substrate on GKE claims 10x compute density and 500 activations per second in a 200,000-agent benchmark.
ElevenLabs details Cue, a forced alignment model that places every character of an audio file on a precise timeline. The company says Cue averages 179 milliseconds of timing error on full recordings, versus 283 to more than 6,000 milliseconds for tested open-source tools, and aligns an hour of audio in about half a second. It covers more than 90 languages.
Anthropic adds a legal artifact gallery to Claude for Legal, offering tools for cite checks, contract review, matter tracking, and regulatory scans, each opening with an example and sourcing links. The company says teams can tailor artifacts with their own data and more than 60 connectors, including Bloomberg Law and Legora, and share them under privacy settings that support ethical walls.
The REA project, an MIT-licensed MCP server that lets AI agents inspect native binaries, Electron apps and .NET assemblies, reports 405,700 GitHub stars and a v6.8.0 release dated Oct. 25, 2026. Setup runs through npx rea-agents setup and requires Node.js 22.x or newer. Deep native analysis depends on Hopper, Ghidra, IDA or Kuna.
Liquid AI releases LFM3, a family of six roughly 2-billion-parameter models spanning text, vision, audio, omni, and decision tasks, built on a 1.7B base. The company says LFM3-1.7B outperforms its LFM2.5-2.6B while using 35% fewer parameters. The open-weight models are available on Hugging Face and Liquid's API, with a 256K-token context window for text.