MONDAY, NOVEMBER 16, 2026 Tech news, minus the noise.
N News Cast
LATEST
Home  /  AI
AI

Kimi K3.1 Pro Open Weights: A 2.8 Trillion Parameter Model You Can Actually Download

Moonshot put the full weights of Kimi K3.1 Pro on Hugging Face on November 2, two days early. At 2.8 trillion parameters with a 1M token context, it is the most capable open model anyone can download right now.

Kimi K3.1 Pro Open Weights: A 2.8 Trillion Parameter Model You Can Actually Download
Source: mikemacmarketing, CC BY 2.0

Moonshot did something unusual on November 2: it kept a promise early. The Chinese lab put the full weights of Kimi K3.1 Pro on Hugging Face two days before the date it had announced, and the download is enormous. The model has 2.8 trillion parameters in total, with 104 billion active at any moment, and the files take up about 1.56 TB. That makes Kimi K3.1 Pro the biggest open model anyone can actually download, and by one respected benchmark, the most capable too.

Kimi K3.1 Pro scores 46.5 on Artificial Analysis's Intelligence Index, which puts it ahead of every other downloadable model. It is not a frontier model in the strict sense. Anthropic's Claude Fable 5.2 sits at 64.6 on the same index, and Google's Gemini 4 Carbon is not far behind. But neither of those can be downloaded, studied, or run on your own hardware. Kimi can, and that difference matters more than the gap in scores.

What Kimi K3.1 Pro actually gives you

The headline spec is the context window: 1 million tokens. That is enough to feed a model an entire codebase, a shelf of legal documents, or a year of company records in one go. Most open models top out far below that, which forces developers to slice documents into chunks and hope nothing important falls between them. A 1M context does not solve every retrieval problem, but it removes the most annoying one.

The price of running it through a hosted provider is already visible. Runware began serving Kimi K3.1 Pro on November 4 at $2 per million input tokens and $10 per million output tokens. That is not the cheapest rate in the market, but it is a fraction of what the frontier labs charge for their best models, and it is the price for a model that also happens to be downloadable.

The catch is the size. At 1.56 TB, Kimi K3.1 Pro is not something you run on a laptop. It needs serious GPU hardware, the kind of setup that costs real money to rent or own. For most people, the practical way to use it is through a hosted provider. The download matters anyway, because it means researchers can inspect the weights, fine-tune the model, and run it without asking a vendor for permission or worrying that an API will change underneath them.

The open weights race is splitting in two

November has been a strange month for open models. On the same week Kimi landed, Tencent shipped Hy4 and Thinking Machines put Inkling 2 Small on its hosted preview while promising full Apache 2.0 weights later. The pattern is that "open" now comes in grades: some models ship weights on day one, others ship a hosted API first and weights later, if ever. Kimi is firmly in the first camp, and that makes it the honest benchmark for what open means in late 2026.

There is also a licensing question worth checking before you build on it. Kimi K3.1 Pro ships under Moonshot's own license, not a standard open source license, so read the terms before using it in a commercial product. The same caution applies to most of this month's releases. The MiniMax H3 family, for example, carries a community license that excludes the US, EU, UK, and South Korea unless you hold a separate license. Open weights do not automatically mean open for business.

The other thing to watch is what the model is actually good at. A 46.5 on the Intelligence Index is a solid general score, but benchmarks are averages, and averages hide weaknesses. Kimi's strength is long-context work and agentic tasks, the kind of jobs where the model has to hold a lot of information in mind and act on it over many steps. If your job is short, sharp reasoning on a single hard problem, a smaller frontier model may still beat it.

Why this release matters beyond the leaderboard

The real story of Kimi K3.1 Pro is not the parameter count. It is that the most capable downloadable model keeps getting better while the frontier labs keep their best models behind applications and approvals. OpenAI's GPT-6 Cyber, released November 5, is available only to approved organizations. Meta's Muse Ember, launched November 9, is cheap and widely available, but it is an API model, not a download.

That split is becoming the defining shape of the AI market. On one side, closed frontier models that you rent by the token. On the other, open models you can own, inspect, and run yourself, with capabilities that now reach into territory that used to belong only to the closed side. Kimi K3.1 Pro does not close the gap, but it narrows it in the place where openness matters most: long documents, private data, and workloads you cannot send to someone else's cloud.

If you run a team that handles sensitive documents or needs a model that works offline, Kimi K3.1 Pro is worth a serious look. Download the weights, test it on your own data, and compare it against the hosted models you already pay for. The benchmark says it is the best open model available. Your own tests will tell you whether that is true for your work, and that is exactly the point of open weights: you get to find out for yourself.

AS
News Cast Editorial Team

We cover AI, gadgets, startups and security — original reporting and explainers, written by humans, for humans. Got a tip? Reach us via the contact page.