Moonshot AI launches Kimi K3.1 Pro to challenge top coding models

Closing the gap on top models

Moonshot AI recently launched Kimi K3.1 Pro, a new model designed to handle complex reasoning and multi-step tasks. As reported by Crypto Briefing, the model is built on the Kimi K3 architecture with specialized training to improve its ability to use tools and solve difficult problems.

The company is positioning this model as a serious competitor to industry leaders like Anthropic and OpenAI. In testing on Terminal Bench 4.0, which evaluates AI agents on real programming tasks, Kimi K3.1 Pro achieved a score of 65.7%. This puts it very close to Anthropic’s Claude Opus 5.5, which scored 66.4%.

The model also outperformed OpenAI’s GPT 6 Astra, which received a score of 58.2% on the same benchmark. Moonshot AI also released a Highspeed variant of the model specifically for users who need responses two to three times faster than the standard version.

Advancements in coding and 3D

One of the primary focuses for Kimi K3.1 Pro is its ability to handle long-horizon coding tasks. These are complex programming jobs that require an AI to perform multiple steps and use various tools over an extended period. On an internal benchmark, the model improved its score from 55.95% to 61.2% compared to its predecessor.

The model shows specific strengths in working with popular game engines like Unity, Unreal Engine, and Godot. It can generate 3D assets, websites, and interactive environments. In a demonstration, the model built a walking simulator using an anatomical model with 296 skeletal components and 402 muscle components.

The developers refined the simulator through 167 tasks and 420 reviews until the virtual character could walk for 11.2 seconds. The model also scored 92.7% on BenchCAD, a 3D modeling evaluation, though it still trails slightly behind Claude Opus 5.5, which scored 96.2%.

Financial and research utility

Moonshot AI is also targeting the financial and business sectors with Kimi K3.1 Pro. The company recorded a 64.4 on OfficeQA Pro v2, a benchmark that measures how well an AI can perform numerical reasoning and retrieve information from financial documents.

Internal testing showed that the new model saw a 7.2 point increase in financial modeling and an 11.6 point jump in financial research compared to the previous Kimi K3. These gains suggest the model is becoming more reliable for investment research and business operations that require high levels of accuracy.

The company also highlighted a significant reduction in hallucinations, which are instances where an AI generates false information. On an internal hard evaluation, the hallucination rate fell from 23.4% to 8.3%. Moonshot noted that while this benchmark tests difficult scenarios, it does not necessarily reflect how the model behaves during everyday use.

Open weights and security

A major milestone for the Kimi K3.1 Pro is the upcoming release of its full weights on November 4. This allows developers to deploy the model independently of proprietary cloud platforms, giving them more control over their infrastructure and data.

Moonshot AI stated that Kimi K3.1 Pro is its most powerful version within the 3 trillion parameter category. To prepare for the public release, the company adjusted its training process to mitigate cybersecurity risks. They removed certain training data and tasks specifically designed to limit offensive cybersecurity capabilities.

The company evaluated the final model on an internal vulnerability research benchmark. The restricted version scored 81% on the test, while a version trained without those specific restrictions scored 90.5%. This move was intended to balance the model's general coding performance with safety requirements for the public.