Ai2 open-sources Olmo-core 3: train trillion-param MoEs 2.7x faster

The open MoE stack hit 858 TFLOP/s on a 1.2T-param run. Also: Microsoft voice models and Voyage reranking reach the gateways; any model gains web browsing.

Nowline OCT 2 4:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A 2.7x-faster open stack for trillion-parameter MoEs

    Ai2 shipped Olmo-core 3 under Apache-2.0 (pip install ai2-olmo-core): a 47B MoE now trains at 52,000 tokens/sec/GPU versus 19,400 before (~2.7x), and it scaled to a 1.2T-param run across 512 NVIDIA B300s at 858 TFLOP/s by keeping experts resident on-GPU instead of resharding (FSDP is out). You get a battle-tested stack to train or fine-tune your own frontier-scale MoEs — if you can rent the GPUs.

  • Microsoft's voice models land on Vercel AI Gateway

    Microsoft's latest speech and transcription models are now callable through Vercel AI Gateway with one API key and unified billing. If you're wiring voice input or audio transcription into an app, that's one fewer provider account and SDK to juggle.

  • Voyage rerank-3 makes RAG reranking cheap

    Voyage's new rerank-3 ($0.05/M tokens) and rerank-3-lite ($0.02/M), both 32K context, are live on OpenRouter. Drop a reranker in front of your retriever to filter irrelevant chunks before they reach the model — now at a price that's hard to argue with.

  • Any model on Vercel can now browse the web

    Vercel AI Gateway added Browserbase Search and Fetch as built-in tools, so any model can discover and pull live web pages via the AI SDK without you standing up a scraper. Weekend build: an agent that researches a topic and cites fresh sources, no custom plumbing.