OVH announces Olala mini 7A1B, its first LLM
Read original source →corporate.ovhcloud.com
The 7B model activates 1B parameters across 256 experts per layer with hybrid Differential-TPA and Mamba-3 attention, running on as little as 8GB of GPU memory at 160 tokens/sec on an RTX A4500. It ships open source on Hugging Face under Apache 2.0 with a 32K context window.
1 source
More stories today
OpenAI launches ChatGPT for Sales with new Sales plugin
OpenAI introduced a Sales plugin plus 30+ new and expanded partner plugins in ChatGPT for researching prospects, preparing pitches, updating CRM, and moving deals forward.
OpenAI Blog·51 minutes ago

Simon Willison releases external sandbox pattern for coding agents
Three sandbox server implementations — external-sandbox-v86 (Alpine Linux in WebAssembly via v86), external-sandbox-container (Apple Containers), and external-sandbox-quicksand (Microsoft's Quicksand) — share one localhost API. Pi, Claude Code, and Codex can be pointed at it to run code outside the agent, with an optional HTTP/HTTPS allow-list proxy and credential replacement.
Simon Willison's Weblog·53 minutes ago

Perplexity adds Multiagent effort level with up to 16 agents
Multiagent sits above Ultra as the highest effort level in Perplexity Computer, with a lead agent assigning work to up to 16 supporting agents across Claude, GPT and Gemini models. Agents share findings directly and cross-check each other's work; Perplexity published no accuracy or speed benchmarks.
MarkTechPost·54 minutes ago

Claude Desktop adds opt-in Latest release channel on Windows and Mac
Claude Developers·1 hour ago
Amazon and Google will together own nearly 30% of Anthropic after IPO
Amazon's stake lands at 14.7% and Google's at 13.7%, both in non-voting shares converted from promissory notes. Amazon is holding back notes that would take it to 23.8%; a group including Nvidia and Microsoft is buying $26B of non-voting stock.
TechCrunch·1 hour ago

Zach Wentz discusses reinforcement learning at Open by Design panel
Together AI·1 hour ago
Vercel's DeepsecBench tests models on unseen codebases
Guillermo Rauch·1 hour ago
Amjad Masad: agent swarms are AI's next scaling frontier
Amjad Masad·1 hour ago