Kimi K3.1 Pro Arrives + NVIDIA’s Small Nemotron Bolt for Local Assistants

AI and SaaS Roundup · 2026-10-25

A new Kimi model for coding and longer agent tasks, plus a much smaller NVIDIA release for local assistants. One runs through a hosted coding service; the other has downloadable weights. Here’s what you can actually use.

Kimi K3.1 Pro is live, but check which plan unlocks it

Kimi rolled out K3.1 Pro in Kimi Code on 23 October, alongside a Highspeed version. The release targets coding, 3D creation and longer agent workflows. Kimi says Highspeed delivers the same model quality at roughly two to three times the speed; that’s its own reported comparison.

The access details are easy to miss: K3.1 Pro starts on Plus / Moderato and above. Pro / Allegretto and above unlock the full million-token context and the Highspeed model. This is a membership access change, not a free credit offer.

There’s also an integration catch. K3.1 Pro only supports thinking mode, so disabling it returns an error. Kimi recommends starting a fresh session rather than switching an existing K3 or older session straight over. Handy details if you’re wiring it into a coding workflow.

Source: Kimi Code’s official release notes

Nemotron Bolt is a smaller option for local text assistants

NVIDIA released Nemotron 3.5 Bolt E4B on 23 October, including a quantized GGUF checkpoint called the 4GB version. It’s post-trained from Gemma 4 E4B and aimed at local device assistants, multi-turn tool use and conversational interfaces.

The model takes text in and produces text out. You could use it as the language component in a voice assistant, but you’d still need speech recognition and text-to-speech around it. A local assistant that calls your app’s tools is another possible use.

The weights use OpenMDW 1.1, and NVIDIA lists commercial and non-commercial use. NVIDIA now includes llama.cpp and Docker examples for running it, with an optional DFlash drafter to try for speculative decoding. The 4GB label also isn’t a guarantee that your entire application will fit in 4GB of memory.

Source: NVIDIA’s official Nemotron Bolt model card