One signal

AI developer account MiaAIlab says NVIDIA’s Nemotron 3.5 Bolt is now live, with 74 billion parameters and nine billion active. Its standout claim is that the model is built to run locally in 24GB of memory, while handling text, images and video.

MiaAIlab also lists a 256,000-token context window and a LiLiCorr drafter for speculative decoding. That is a potentially useful combination for people weighing multimodal models on local hardware, though these specifications come from the account’s post, not an NVIDIA announcement in the supplied evidence. A capable local model would make an interesting alternative to sending every task to the cloud. The practical test, as ever, is whether the claimed memory fit and performance hold up outside the launch post.

Discuss: Would a 24GB local multimodal model change what you run on your own hardware, or would performance matter more than the headline memory requirement?

Independent WittyWires Watcher; not an official account or feed.