2.8 trillion parameters: 75% more than DeepSeek

Simply put, parameters are the model's "neural connections" — the more there are, the more knowledge and patterns the model can absorb. Kimi K3 has 2.8 trillion, which is roughly 75% more than the previous largest open model, DeepSeek V4 Pro (about 1.6 trillion). Moonshot AI thus takes the symbolic crown of the largest open model, which Chinese companies have held for some time — DeepSeek is followed by Xiaomi (1.02 trillion) and Alibaba (397 billion).

An important detail: K3 is a so-called MoE model (Mixture of Experts). This means that not all parameters work at once for each query — the model activates only 16 out of 896 specialized "experts." Thanks to this, operation is significantly cheaper than anyone would expect from a model of this size.

Add to that a context window of 1 million tokens (the model can process the content of roughly ten books at once), native image and video processing, and a permanently enabled "thinking" mode where the model reasons through its approach before answering. Details are described on the official Moonshot AI blog.

Benchmarks: Third in the world, often first in agent tasks

According to data from analytics firm Artificial Analysis, which independently evaluated the model, K3 placed third on the GDPval-AA v2 benchmark (measuring real-world work tasks across 44 professions) with a score of 1,687 — ahead of it are only Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8). It did, however, beat Claude Opus 4.8 (1,600), for example.

Even more interesting are agent tasks, i.e., situations where the AI autonomously works on a long task. On the BrowseComp benchmark, which tests complex information retrieval, K3 achieved a record 91.2 points — the best result of all models, including closed ones. In the Frontend Code Arena ranking, where people rate the quality of generated web code, it took first place ahead of both Claude and GPT. According to VentureBeat, it won four out of eight real-world work automation tests.

Moonshot itself admits that overall, K3 still slightly lags behind Claude Fable 5 and GPT-5.6 Sol — for instance, in the challenging academic test HLE-Full it scores 43.5 points against Fable 5's 53.3 points. But the open model is no longer playing in a different league; it's more of a draw with the top tier.

A chip designed in 48 hours without human input

The most impressive demonstration is an experiment in which K3 was tasked with designing a physical chip that would run a miniature version of itself. The model worked continuously for 48 hours as an autonomous agent: reading documentation, making architectural decisions, running verifications, and fixing its own bugs in open-source electronic design tools. The result is a chip design with an area of 4 mm², which ran at 100 MHz in simulation and decoded over 8,700 tokens per second.

Similarly, the model managed to reproduce a complex astrophysical calculation (the universal I-Love-Q relation for neutron stars), which would take an experienced scientist one to two weeks — K3 read over 20 scientific papers and wrote a complete computational pipeline in about two hours.

The comeback of a company DeepSeek nearly buried

Kimi K3 is the culmination of a remarkable turnaround. Moonshot AI was founded in 2023 by Yang Zhilin, a Tsinghua University graduate with a research background at Google and Meta. The Alibaba-backed company was among the stars of the Chinese AI scene — but the rise of cheap DeepSeek R1 in January 2025 hit it hard: the Kimi app dropped from third to seventh place in Chinese traffic rankings.

The response was a shift to open models: Kimi K2 in July 2025, K2.5 this January, and now K3. The company, valued at 4.3 billion dollars according to Reuters, timed the release just before the global AI conference WAIC in Shanghai — sending a clear message that it wants to be the center of the global open-source community.

Price, availability, and what it means for Czechia

You can try the model for free right now at kimi.com — just sign in with a Google account or phone number; no payment card is required. The interface also communicates in Czech, though the primary languages remain English and Chinese. Via API, K3 costs 3 dollars per million input tokens and 15 dollars per million output tokens; with cached repeat queries, the input price drops to 0.30 dollars. That's on par with mid-range models from Western labs — for performance just below the absolute top tier.

For European and Czech companies, however, the key date is July 27, when Moonshot will release the full weights. An open model of this class can be downloaded, modified, and run on one's own infrastructure — meaning data doesn't have to leave the company or the EU, which simplifies compliance with GDPR and the European AI Act. The catch is the hardware requirements: inference for a model with 2.8 trillion parameters requires dozens of accelerators; Moonshot recommends configurations with 64 or more GPUs. For most Czech companies, the more realistic path will be via API or the smaller K2.6 and K2.7 Code models (both at 0.95/4 dollars).

One thing is certain: the era when open models lagged six months behind closed ones has just ended. And pressure on GPT and Claude pricing will only intensify.