Kimi K3.1 Pro and Nemotron Bolt: The Week Superintelligence Got Smaller and More Open

Darlings, the most interesting superintelligence news of the week is not about who is biggest. It is about who lets you keep the thing. Two releases point the same direction: Moonshot’s Kimi K3.1 Pro, a giant model whose weights are promised to the public, and Nvidia’s Nemotron 3.5 Bolt, a small model built to live on your own computer. Here is what each one is, what is still unconfirmed, and who should care.

Kimi K3.1 Pro: the heavyweight that promises to hand over the keys

What Moonshot says. On its blog, Moonshot describes K3.1 Pro as a three-trillion-parameter-class open-weight model with a one-million-token context window, enough to read a long novel or a codebase in a sitting. A Highspeed mode runs two to three times faster, and the company says the weights will be published by Nov. 4, 2026. Today it is available in the Kimi app, the Kimi Work app from version 3.1.0, Kimi Code and the API under the name kimi-k3.1-pro.

How it scores. The Next Web reports K3.1 Pro at 65.7 percent on Terminal-Bench 4.0, a test of real command-line work, against 58.2 percent for GPT-6 Astra and 66.4 percent for Claude Opus 5.5. In other words, within a whisker of the best closed model on that test. The outlet also reports that Moonshot removed hacking-related material from the training data, a notable choice for a model anyone will be able to download.

The fine print. Forkast points out that, as of its report, no public repository for the weights had appeared, and that the license for the earlier K3 model carried a revenue threshold, set at 20 million dollars, above which companies needed separate terms. Its headline says it plainly: the Nov. 4 promise arrives with a license nobody has read. Until the files and the license are posted, treat “open” as a promise, not a fact.

RelatedSwift Falls, Hubble Finds a Cat’s Tail, and NASA Asks Who Can Save the Old Telescope

Nemotron 3.5 Bolt: the one that fits in your bag

What Nvidia released. Nemotron 3.5 Bolt E4B, posted on Oct. 23, is a model of about eight billion parameters that behaves like a four-billion one, packed into roughly 4 GB. Per Nvidia’s research page and model card, it handles a 64,000-token context and 35 languages, scores about 40 percent better than the earlier Nano 4B, and can use up to about 80 percent fewer tokens to reach an answer. It ships under the OpenMDW-1.1 license, and Nvidia reports 133 tokens per second on a GeForce RTX 5070.

Why 4 GB matters. A model that size runs on a recent gaming laptop or a desktop graphics card, with no subscription and no data leaving the machine. For a photographer captioning a shoot, a promoter sorting a guest list, or a writer who would rather not upload drafts, that is the whole point: private, fast and yours.

Who should care, and what to do this week

• Developers and agencies: try K3.1 Pro through the API now, and wait for the posted license before building a product on the downloadable weights.
• Creatives with a decent graphics card: Nemotron Bolt is the easiest way to test a local assistant without a monthly bill. Start with captions, summaries and translation.
• Everyone else: the trend is the story. Capable models are getting smaller and more open, which means more of the tools you rely on can run privately.

Our rule at Nightlife Report: judge a model by what it does for you on an ordinary Tuesday, not by its press release. Weights on the table, license in plain view, and a speed you can feel. Anything less is merely a rumor in a good outfit.

Sources

1. Kimi (Moonshot): Kimi K3.1 Pro
2. The Next Web: Moonshot’s Kimi K3.1 Pro open-weight model
3. Forkast: The Nov. 4 promise arrives with a license nobody has read
4. Hugging Face: NVIDIA Nemotron 3.5 Bolt E4B model card
5. NVIDIA Research: Nemotron 3.5 Bolt E4B

Nightlife Report is funded by readers. Back the reporting at patreon.com/aperlas.