Level 1 — Absolute Beginner
A company in China called Moonshot AI made a new computer program. It is called Kimi K3. Kimi K3 can understand pictures and words, and it can also write computer code.
This week, Moonshot AI gave away the full 'weights' of Kimi K3. Weights are like the brain of an AI program. When a company gives away the weights for free, other people can use and study the AI program.
Kimi K3 is very good at writing computer code. On some tests, it did better than many other AI programs from other companies. People call this kind of test a benchmark.
Many companies in the United States and China are racing to build the best AI programs. Kimi K3 shows that Chinese companies are building AI that can compete with AI from other countries.
- company
- a business that makes products or provides services
- computer program
- a set of instructions that tells a computer what to do
- weights
- the numbers inside an AI program that control how it thinks and responds
- free
- costing no money
- code
- instructions written for a computer, used to build programs and websites
- test
- an activity used to check how good something is
- benchmark
- a test used to compare how well different programs perform
- compete
- to try to do better than others in a contest
Level 2 — Elementary
The Chinese AI company Moonshot AI has released the complete open weights of its newest model, Kimi K3, making it one of the largest open-weight AI models in the world. Open weights means the underlying files that power the model are shared publicly, letting developers everywhere download, study, and build on it.
Kimi K3 uses a design called a Mixture-of-Experts architecture, which contains 2.8 trillion parameters in total, spread across 896 smaller expert networks. For any given task, only 16 of those experts activate at once, which helps the model run efficiently despite its enormous size. It also supports a one-million-token context window, letting it process very long documents at once.
On coding benchmarks, Kimi K3 has performed impressively, ranking first on tests including Program Bench, SWE Marathon, and the Frontend Code Arena, where it scored an Elo rating of 1,679. It still trails top closed models from Anthropic and OpenAI on some overall evaluations, but its performance has surprised many industry analysts.
The release adds to a wave of powerful open-weight models coming out of China this year, intensifying competition with US AI labs and giving developers around the world free access to frontier-level AI technology outside the traditionally closed American ecosystem.
- open weights
- AI model files that are shared publicly for anyone to download and use
- architecture
- the overall design or structure of a computer system or AI model
- parameters
- internal values in an AI model that are adjusted during training
- expert networks
- smaller specialized parts within a larger AI model, each handling certain tasks
- context window
- the amount of text an AI model can process or 'remember' at one time
- Elo rating
- a scoring system used to rank competitors based on head-to-head performance
- frontier-level
- at the cutting edge of what current technology can achieve
- ecosystem
- a network of companies, tools, and technologies that work together in an industry
Level 3 — Intermediate
Moonshot AI, the Chinese startup behind the Kimi chatbot, has released the complete open weights of its latest model, Kimi K3, positioning it as one of the largest open-weight AI systems ever made publicly available. Unlike closed models offered only through paid APIs, open weights allow developers worldwide to download the model's underlying parameters directly, inspect its internals, and fine-tune it for specialized applications.
Architecturally, Kimi K3 is built as a sparse Mixture-of-Experts system comprising 2.8 trillion total parameters distributed across 896 expert modules, of which only 16 activate per token processed, an approach that balances enormous model capacity against practical computational cost. The model also supports a one-million-token context window and was trained in part using clusters of Nvidia's GB300 chips.
On coding-focused benchmarks, Kimi K3 has posted standout results, topping evaluations such as Program Bench, SWE Marathon, BrowseComp, SpreadsheetBench 2, and Automation Bench, and claiming the top spot on the Frontend Code Arena with an Elo rating of 1,679. It nonetheless trails Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on select overall evaluations, underscoring that the frontier remains contested rather than settled.
The release lands amid a broader surge of open-weight models emerging from Chinese labs this year, a trend reshaping competitive dynamics in the AI industry by giving enterprises and independent developers access to near-frontier capability outside the tightly controlled ecosystems maintained by leading American vendors.
- startup
- a newly established, often small, business or company
- sparse
- in AI, a design where only a portion of a model's components are active at once
- modules
- self-contained components that make up part of a larger system
- computational cost
- the amount of processing power and resources required to run a task
- clusters
- groups of connected computers or processors working together
- standout
- notably impressive or better than others in comparison
- contested
- disputed or actively competed over
- surge
- a sudden, powerful increase in activity or quantity
Level 4 — Advanced
Moonshot AI, the Chinese startup behind the Kimi line of chatbots, has released the full open weights of its newest model, Kimi K3, establishing it among the largest open-weight AI systems made publicly available to date. Unlike proprietary systems accessible only through metered commercial APIs, open weights afford developers worldwide unmediated access to a model's underlying parameters, enabling direct inspection, modification, and fine-tuning for specialized downstream applications, a transparency that closed labs have historically declined to offer.
Architecturally, Kimi K3 is instantiated as a sparse Mixture-of-Experts system comprising 2.8 trillion total parameters distributed across 896 discrete expert modules, of which a mere 16 activate per processed token, a design that reconciles vast representational capacity with tractable inference-time computational overhead. The model further supports a one-million-token context window and was trained in part on clusters of Nvidia's GB300 accelerators, underscoring the immense hardware investment underpinning frontier-scale model development.
On coding-oriented benchmarks, Kimi K3 has registered standout performance, claiming top rankings across Program Bench, SWE Marathon, BrowseComp, SpreadsheetBench 2, and Automation Bench, and securing the leading position on the Frontend Code Arena with an Elo rating of 1,679. Nonetheless, it continues to trail Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol on select composite evaluations, a reminder that even exceptional open-weight releases have not yet fully closed the gap with the most capable proprietary systems.
The release arrives amid an accelerating cascade of open-weight releases from Chinese AI laboratories this year, a development with consequential implications for the competitive geometry of the global AI industry, as it extends near-frontier capability to enterprises and independent developers operating outside the tightly governed ecosystems long maintained by dominant American vendors.
- proprietary
- privately owned and controlled, with restricted access for others
- unmediated
- occurring directly, without an intervening party or process
- instantiated
- represented as a concrete, specific instance of a more general concept
- representational capacity
- the extent to which a system can encode and express complex information
- tractable
- manageable or able to be handled or dealt with practically
- accelerators
- specialized hardware components designed to speed up specific computational tasks
- composite evaluations
- assessments that combine multiple individual tests into a single overall measure
- geometry
- in this context, the overall structure and arrangement of relationships within a field
Talk about this article with KatePractice this story's vocabulary in a real spoken conversation with your AI tutor