๐—š๐—ผ๐—ผ๐—ด๐—น๐—ฒ ๐—ฅ๐—ฒ๐—น๐—ฒ๐—ฎ๐˜€๐—ฒ๐˜€ ๐——๐—ถ๐—ณ๐—ณ๐˜‚๐˜€๐—ถ๐—ผ๐—ป๐—š๐—ฒ๐—บ๐—บ๐—ฎ

Google released DiffusionGemma. It has 26 billion parameters. It uses diffusion to create text. This process turns noise into words.

Nvidia says it hits 1,000 tokens per second on one H100 GPU. It runs 4 times faster than Gemma 4.

Key facts:

It produces lower quality output. Google calls it an experimental tool for developers. Use it for on-device tasks. Use it for real-time needs.

Find it on Hugging Face. Nvidia hosts free inference on the NIM cloud API.

Source: https://dev.to/gentic_news/google-open-sources-diffusiongemma-26b-model-hits-1k-tokenssec-on-h100-2nak Optional learning community: https://t.me/GyaanSetuAi