Granite Switch 4.1: chat, decisions, and speech in one open model
Two open checkpoints fold a transcribe-decide-respond loop onto one GPU. Plus NVIDIA's open 2.8T-token Indic corpus and Gemini 4 Argon's price cliff.
Copy markdown
One checkpoint, three jobs
IBM's Granite Switch 4.1 (8B and 30B, open weights on Hugging Face) wraps a base chat model with Activated-LoRA adapters for routing decisions, safety guarding, and Granite-Speech transcription. Because the adapters reuse the base model's KV cache, the whole stack fits in roughly the VRAM of the base model alone — so one endpoint takes a spoken request, makes a routing decision, and returns a natural-language reply.
Build this weekend: a local voice agent
IBM ships a Speech-to-Decisions demo that transcribes audio, summarizes it, screens for harmful content, and checks your own custom flags — a full voice-to-action loop without chaining three separate models or paying per-call API fees. You can also compose your own Switch model from Granite Libraries adapters.
Gemini 4 Argon's API price doubles Dec 1
If you're building on Google's new frontier model, watch the clock: introductory API pricing of $2/$10 per million input/output tokens runs only through Nov 30, after which it reportedly jumps to $4/$20. The 1M-token model — with code execution, function calling, and file search — is live now in the Gemini API and AI Studio.
NVIDIA opens a 2.8T-token Indic corpus
Nemotron-pretraining-Indic-v1 brings ~2.8 trillion tokens across 22 Indian languages, plus the GPU-accelerated generation pipeline (prompts, workflows, quality checks) in NeMo Data Designer. Stream samples by language without downloading the full set, or point the pipeline at your own documents to build domain data in low-resource languages.