Moonshot AI launches the new flagship model Kim K3.1 Pro with 3T parameters

Moonshot AI today introduced Kim K3.1 Pro, a large language model that can perform some tasks better than Claude Opus 5.5 and GPT-6 Astra.

The LLM is a retrained or updated version of an algorithm that the Chinese artificial intelligence company released in July. The latter model, Kimi K3, includes 896 individual neural networks with 3 trillion parameters. Only 16 of these neural networks are activated at a time to minimize inference costs.

Moonshot AI also optimizes the hardware footprint of its models in other ways. Kim K3 features Kimi Delta Attention, a custom implementation of a memory-saving technology called Gated DeltaNet. This reduces the amount of RAM that the model’s attention mechanism uses to interpret user prompts. The function has probably been transferred to Kim K3.1 Pro.

The company says the new model’s post-training workflow used significantly more computing power than similar projects it has undertaken in the past. Additionally, it ran on an updated version of the company’s post-training stack. Moonshot AI added automated quality checks and increased the efficiency of the underlying infrastructure.

One area where Kimi K3.1 Pro delivers particularly competitive output quality is 3D modeling. There are two ways to create a 3D model of an object: engineers can draw it or write a code snippet that numerically specifies its geometry. Kimi K3.1 Pro takes the latter approach. The LLM scored 92.7% on a 3D model coding benchmark called BenchCAD, putting it within four percentage points of Astra and Claude Opus 5.5.

Kimi K3.1 Pro can not only generate static 3D models, but also animate them. In one test, Moonshot AI used the model to generate an animated simulation of human movement. Another internal evaluation revealed that the company was developing several 3D video games. It is difficult for AI models to generate games because they contain a large amount of code and even a single mistake can cause them to crash.

Pro achieved a score of 65.7% on the industry-standard Terminal Bench 4.0 coding benchmark, putting it more than 7% ahead of Astra but slightly behind Opus 5.5. Additionally, it managed to outperform both models on OfficeQA Pro v2. This is a test that measures the ability of LLMs to perform complex financial analysis tasks.

The model’s performance on popular cybersecurity benchmarks lags significantly behind Astra and Opus 5.5. Moonshot AI says its engineers intentionally removed cybersecurity information from the model’s post-training dataset to address security risks. The company plans to add advanced vulnerability analysis capabilities to its future models.

Kimi K3.1 Pro will be available through Moonshot AI’s paid cloud services at launch. The company plans to open source the model’s weights by November 4th.

Image: Unsplash