Moonshot AI to release Kimi K3.1 Pro weights on November 4

A new frontier for local AI

Research groups and organizations that require high-level AI performance often face a dilemma regarding data privacy. Many prefer to keep sensitive information off the servers of major vendors, but this usually limits them to using less capable models. Moonshot AI is addressing this by preparing to release the full weights for its Kimi K3.1 Pro model on November 4.

By releasing the weights, Moonshot allows developers to host the model on their own hardware. This means a company could run a near-frontier model locally, ensuring that data remains within their own infrastructure. This move is significant for users who need the reasoning power of a top-tier model without the risks associated with cloud-based APIs.

Independent testing by Artificial Analysis shows that K3.1 Pro is already competitive with the best models currently available. It earned a 55 on the Intelligence Index, which ranks it fourth out of 215 different models. This puts it in the same league as major releases from companies like Anthropic and Google.

Comparing performance and costs

The pricing for Kimi K3.1 Pro has become increasingly competitive with Western counterparts. Moonshot currently charges $2 per million input tokens and $10 per million output tokens, which matches the rates for OpenAI's GPT-6.1 Sol. These prices are significantly lower than the $5.98 per task cost for Claude Opus 5.5, as calculated by Artificial Analysis.

While the cost is lower, the performance varies depending on the specific task. In the SWE Marathon, a test for long-horizon software engineering, K3.1 Pro scored 51.7, beating Claude Opus 5.5's score of 46.3. However, it struggled in scientific research workflows, scoring 38.1 on the Terminal-Bench Science 0.1 test compared to 68.1 for GPT-6 Astra.

Moonshot also offers a Highspeed version of the model for users who prioritize speed over cost. This variant runs two to three times faster than the standard version but costs $4 per million input tokens and $20 per million output tokens. This higher-priced option matches the speed of Opus 5.5 but costs twice as much per token as the lower-cost GPT-6.1 Sol.

Technical hurdles and successes

Moonshot AI demonstrated the model's capabilities through a simulation involving a human-atlas, which is an open-source anatomical model. The goal was to create a walking simulator where a digital human could move realistically. During the process, the model encountered issues with the movement of the arms and had to investigate the simulator itself to find errors.

The model successfully identified and corrected 36 calculation errors regarding how shoulder muscles rotate the joints. After retraining the walking controller with these corrections, the simulated human was able to perform alternating strides for 11.2 seconds. This demonstrates the model's ability to perform complex reasoning to solve problems within a simulated environment.

Despite these successes, some experts remain cautious about its standing against the absolute best models. Peter Gostev, an AI capability lead at Arena, noted in a video that the model still lags behind Western rivals in certain tasks. He specifically highlighted its performance in creating 3D scenes, animations, and websites, stating, "At least in these tests, I don’t think it’s quite there."

The shifting AI landscape

The release of Kimi K3.1 Pro highlights a shift in the global AI market where Chinese models are no longer significantly undercutting Western ones. Instead, the pricing and performance of leading models from both regions are beginning to overlap. This convergence makes it easier for developers to choose the best tool for a specific job based on cost-efficiency rather than just geographic origin.

For researchers, the upcoming release of the weights on November 4 is the most important detail. It provides a bridge between high-end performance and the need for private, local execution. As more models reach near-frontier levels, the ability to run them on-premises will become a standard requirement for many industries, including finance, law, and health.

Moving forward, the industry will watch how K3.1 Pro performs in real-world deployments compared to its rivals. While it excels in some software engineering benchmarks, its lower scores in scientific workflows suggest there is still room for improvement. The next few months will determine if Moonshot can maintain its momentum as it moves toward making its most capable models more accessible to the public.