Here's a short video from our founder, Zhilin Yang.

Hi, I'm Jayen, the founder of Kimmy. Six

months ago, we released Kimmy K2.

[music] It was the first open model to scale for training to one trillion parameters. Later, we introduced Kimmy

parameters. Later, we introduced Kimmy K2 thinking it scale the thinking time of agents for solving long horizon tasks. It performed interle thinking

tasks. It performed interle thinking with two calling for up to 300 steps all on its own. Today, we're dropping our latest model Kimmy K2.5. K2.5 is a

powerful model that excels at [music] agents coding vision and general capabilities and it is open source. K2.5

[music] sets new records on the most challenging agentic benchmarks including humanities last exam browse comp and

deep search QA. K2.5 is also a very strong calling model.

We didn't just want key me to code. We

wanted it to have an eye for [music] design. It gives designer grade websites

design. It gives designer grade websites a flow with grace and visual poetry.

K2.5 lowers the barrier of entry with the vision capabilities. Simply upload a screen recording. Kim K2.5 will rebuild

screen recording. Kim K2.5 will rebuild it from scratch with clean professional code. Going beyond front end design,

code. Going beyond front end design, K2.5 is great as software engineering.

Kim Code is our new coding product. Kim

code supports images and videos as inputs and also automatically discovers existing skills into your working environment in Kimico. Next, office to

make Kimmy even more helpful in your daily workflow. K2.5 has mastered the

daily workflow. K2.5 has mastered the core skills of the office [music] suite.

Whether it is building a complex financial model, handling the formatting of a professional PDF or crafting a consulting level deck, we want to give

that power to everyone. Tasks that used to take days like merging 50 different department reports or turning a 30,000 word paper into a precise pitch deck now

happen in 10 minutes. With K2.5, we go from one agent to an agent's one.

Instead of asking one agent to do everything, K2.5 creates and coordinates each realm of specialized agents to work in parallel. This specialized agents are

in parallel. This specialized agents are essentially copies of K2.5, but with different rows and different subtask.

Nothing is predefined. The rows and subtasks are created on the fly by K2.5.

Running in parallel substantially reduces the time needed for a complex task. For market research spanning 100

task. For market research spanning 100 companies, it orchestrates a swarm of analysts for turnar around in minutes not weeks. For complex translation

not weeks. For complex translation projects of 300 pages, it mobilizes a team of linguists for [music] rapid accurate delivery. For literature

accurate delivery. For literature synthesis across 50 papers, it builds 10 senior researchers to conduct [music] parallel analysis and generate a

comprehensive survey instantly. Training

the agent swarms at scale is technically challenging. We rebuilt our

challenging. We rebuilt our reinforcement learning infrastructure and optimize the training algorithms to ensure the best efficiency and performance. All of these capabilities

performance. All of these capabilities come together in Kim K25, a powerful open-source model. You can experience

open-source model. You can experience all of this starting today on kim.com or the Kimmy app. For software engineers, we recommend pairing K2.5 with Kimmy

code and also K2.5 is available via our API. Go play with it and tell us what

API. Go play with it and tell us what you think. Thanks.

you think. Thanks.