Moonshot Launches Kimi K3.1 Pro, Open Weights Promised by Nov 4

October 25, 2026 · Cortex Solutions

Moonshot AI released Kimi K3.1 Pro, a 1M-context model it says rivals top closed models, with weights due by Nov 4. What it means for Saudi businesses.

Chinese company Moonshot AI this week launched Kimi K3.1 Pro, the first model in its new K3.1 series, and promised to make its full weights free to download by November 4, 2026. The announcement comes as competition between open-weight and closed models intensifies, and it is worth watching for organizations looking for AI options they can run on their own servers.

This article relies on two sources published on October 23, 2026: Unite.AI's report on the launch, and a Startup Fortune report by Judith Murphy. All figures here were reported by the company itself and have not yet been independently verified.

What is Kimi K3.1 Pro?

According to Unite.AI, K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute, and the company calls it its "most capable 3T-class open-weight model to date." Startup Fortune says it is a mixture-of-experts model with about 2.8 trillion total parameters, of which only about 104 billion are active per token, which makes it lighter to run than its total size suggests.

Key features include:

• A context window of up to one million tokens, so it can read very large documents, contracts and code repositories in one pass.
• A faster variant, Kimi K3.1 Pro Highspeed, running about two to three times faster with the same capabilities, according to Unite.AI.
• It always runs in thinking mode and accepts image and video input, according to the developer docs cited by Unite.AI.
• Zero Data Retention is available for API customers.

What do the numbers say?

Both sources relayed the company's own figures, including:

• On Terminal-Bench 4.0, for command-line coding tasks, the model scored 65.7%, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra.
• On BenchCAD with tools it scored 92.7%, below Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%.
• On an internal long-horizon coding benchmark it rose to 61.2% from 55.95% for the previous K3.
• On an internal hallucination benchmark, the claim-level error rate fell from 23.4% to 8.3%, with the company noting, as Unite.AI reports, that these rates do not reflect overall error rates in daily use.

Unite.AI also reports gains on internal finance benchmarks, such as financial modeling and financial research, compared with the previous model.

The picture, then, is a model that comes close to the strongest closed models on some tasks and trails them on others, with the final comparison awaiting independent testing.

Open weights and cyber restrictions

For now the model is available only through Kimi's apps, website and API, and cannot yet be self-hosted. But the company has promised, as Startup Fortune reports, that "the full weights are free to download by November 4." Neither source states the license under which they will be published.

The notable new element is that Moonshot removed cyber-related data from both training stages of the open-weight version, while keeping general coding ability. According to Unite.AI, a checkpoint without these measures scored 90.5% on an internal exploit benchmark, while the final model scored 81.0%. The company says its cyber capability remains behind frontier closed models that have been public for at least three months, and that it will not release open weights that extend the absolute cyber frontier.

This matters because open models, once published, cannot be withdrawn or controlled, so the company is trying to balance openness with limiting misuse.

How can it be accessed now?

According to Unite.AI, the model is available on kimi.ai and the mobile apps, in the Kimi Work desktop app, in the Kimi Code tool for developers, and via the API under the IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Using it requires a Plus plan or above, while the one-million-token window and the Highspeed variant require Pro or above. For Kimi Enterprise customers, access is off by default and can be enabled by an admin.

What does this mean for Saudi businesses?

• A new option for local deployment: once the weights are published, organizations that require their data to stay on their own servers or inside the Kingdom will be able to test a strong model without sending data to an external provider, bearing in mind that the model's size requires substantial compute.
• Don't rely on the company's numbers alone: the reported figures are self-reported, so it is best to test the model on your actual tasks, in Arabic in particular, before any decision.
• Check the license before commercial use: the license has not been announced yet, and its terms will determine whether it suits your products and services.
• Governance first: check data protection policies and the requirements of Saudi regulators before passing customer data to any model, open or closed.
• Competition benefits users: open models closing the gap with closed ones means more choice, lower cost, and a chance to build AI solutions tailored to the local market.

At Cortex Solutions we help organizations choose the right model and build AI agents and automation that work securely on their own data. Learn about our services: https://cortex.sa/services/ai-automation