Chinese artificial intelligence developer Moonshot AI today debuted Kimi K2.5, an open-source model that it says can outperform GPT-5.2 across several benchmarks.

The launch comes a few days after word emerged that the company is raising capital at a $4.8 billion valuation. Moonshot reportedly closed a separate $500 million round in December.

Kimi K2.5 is derived from a large language model called Kimi K2-Base that the company released in early November. One of the latter model’s flagship features is that it uses an algorithm called Muon to speed up training. Muon boosts performance by accelerating an LLM’s hidden layers, the modules that perform the bulk of the calculations involved in answering prompts. 

According to Moonshot, its engineers enhanced Kimi K2-Base by training it on 15 trillion tokens’ worth of data. The dataset included not only text but also multimodal files. As a result, Kimi K2.5 is better than its predecessor at processing multimodal files such as charts.

Moonshot says that the model features a mixture-of-experts architecture with 1 trillion parameters. Those parameters are organized into multiple neural networks that are each optimized for a different set of tasks. When Kimi K2.5 receives a prompt, it doesn’t activate all its parameters but only the specific neural network that is best equipped to generate an answer. The result is a significant reduction in hardware usage.

The neural networks that make up Kimi K2.5 each include about 32 billion parameters. They’re supported by a so-called vision encoder with 400 million parameters. According to Kimi, it’s responsible for translating multimodal data uploaded by users into embeddings. Those are abstract mathematical representations that Kimi K2.5’s artificial neurons can understand.

LLMs use a mechanism called attention to review the data at their disposal and find the details that are most relevant to the task at hand. According to Moonshot, Kimi K2.5 parallelizes the calculations that its attention mechanism uses to identify relevant details. That approach boosts performance because performing calculations side-by-side is faster than completing them one after another.

Kimi has a standard mode and a Thinking mode that offers higher output quality. Additionally, a capability called K2.5 Agent Swarm enables the LLM to split complex tasks into simpler sub-steps and assign each sub-step to a separate AI agent. A built-in orchestration engine can create and manage up to 100 agents per prompt.

The Kimi K2.5 Agent Swarm serves a similar purpose as the model’s parallelized attention mechanism. Agents can perform sub-steps concurrently instead of one after one another to reduce waiting times.

Moonshot compared Kimi K2.5 against GPT-5.2, Claude 4.5 Opus and other reasoning models across more than two dozen benchmarks. The company says that its model achieved the highest score on HLE-Full, one of the industry’s most difficult LLM evaluations. It comprises 2,500 questions spanning fields such as math and physics.

In most of the other benchmarks, Kimi K2.5 came within a few percentage points of the other LLMs’ scores. It bested GPT-5.2 on several occasions. 

Moonshot has made the Kimi K2.5’s code available on Hugging Face.

Image: Unsplash