The first open 3T-class model

Kimi K3, which Moonshot AI introduced on July 16, pushes the size boundary of open models to 2.8 trillion parameters — roughly double the previous generation Kimi K2.6. This is not brute force without thought: the model uses a Mixture of Experts architecture, where only 16 out of 896 experts are activated per query, i.e., a fraction of the total capacity. As a result, the model runs significantly cheaper than its paper size would suggest.

The technical foundation rests on two architectural innovations. Kimi Delta Attention (KDA), according to Moonshot, accelerates decoding in million-token contexts by up to 6.3× — in practice, this means the model can work nimbly even with entire document libraries or large code repositories. The second innovation, Attention Residuals, brings approximately 25% higher training efficiency at less than 2% additional cost. Combined, Moonshot reports roughly 2.5× better scaling efficiency compared to Kimi K2 — meaning more intelligence from every unit of compute invested.

Dethroned Claude in frontend, but still catching up overall

The loudest response came from the independent Arena.ai. Kimi K3 climbed to first place in the Frontend Code Arena ranking with a score of 1,679 points, overtaking Claude Fable 5 — a leap of 17 places compared to Kimi K2.6, which ranked only 18th. In six of the seven frontend domains (marketing, design from reference, data applications, consumer products, simulations, and content creation tools) it holds first place; only in the games category does it remain second behind Fable 5. In direct output comparison, Kimi K3 won in 76% of pairwise battles, while Claude Fable 5 reaches 63% and GPT-5.6 Sol 58%.

Moonshot itself, however, is unusually sober in its own communication: it admits that in overall performance, Kimi K3 still lags behind the strongest proprietary models Claude Fable 5 and GPT 5.6 Sol. The numbers confirm this — on the agentic ranking GDPval-AA v2, K3 scores 1,687, good enough for third place behind Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,748). On the GPQA Diamond test (PhD-level science questions) it reaches 93.5%, on Terminal-Bench 2.1 for command-line work 88.3%, and in the BrowseComp benchmark for web research it holds the best published score at 91.2%.

A chip designed by the model, a compiler in a few days

Moonshot accompanied the release with a series of case studies focused on long autonomous tasks. For instance, during a single 48-hour autonomous run, Kimi K3 designed, optimized, and verified a chip using open-source EDA tools — the resulting 4mm² design handles over 8,700 tokens per second in simulation. In another test, the model built MiniTriton from scratch, a simplified GPU compiler whose performance in supported workloads rivals the established Triton. And in a scientific scenario, it reproduced a computational astrophysics study in two hours that would take an experienced researcher one to two weeks — including cross-validation of over 20 scientific papers.

The entire product suite targets similar tasks: Kimi Work (a desktop app for knowledge work with new widgets and dashboards), Kimi Code (a terminal agent), and mobile apps for iOS, Android, and HarmonyOS.

Sonnet-level pricing — and a competitor stock crash

Through the API, Kimi K3 costs $3 per million input tokens and $15 per million output tokens (with cache hits only $0.30 for input). That is the same price tier as Claude Sonnet and simultaneously a significant jump from Kimi K2.6, which cost $0.95/$4 — making Kimi K3 the most expensive model a Chinese AI lab has ever offered. Yet it is still several times cheaper than Claude Fable 5 ($10/$50).

That this is a significant moment was also signaled by the stock market: according to agency reports, shares of rival Chinese lab Z.ai plunged 28% after the announcement, and MiniMax Group lost 16%. Commentator Simon Willison, in his analysis, nevertheless tempers the enthusiasm: he notes that the model currently offers only a single reasoning mode ("max") and that the decisive factor will be reliability during long agentic conversations, which benchmarks measure only partially.

What it means for Czechia and Europe

Kimi K3 has been available without restrictions from Czechia since July 16 — via the web at kimi.com, mobile apps, and the API at platform.kimi.ai. As a multilingual model, it handles Czech as well, though Moonshot does not offer an official Czech localization of the interface. More significant is the promise of open weights by July 27: European companies and research institutions will be able to run the model on their own infrastructure, which is crucial for sectors with strict data protection requirements. The catch is the size — Moonshot recommends deployment on setups with 64 or more accelerators, so self-hosted operation will remain the domain of larger players and cloud providers. For EU deployment, as with all large models, the AI Act and its obligations for general-purpose AI models, which have been in effect since August 2025, will also play a role.

Moonshot also openly admits weaknesses: the model is sensitive to correct handling of reasoning history (the maker recommends proven tools like Kimi Code), it may act too autonomously with unclear prompts, and in overall user experience, by the company's own words, it "noticeably lags" behind Claude Fable 5 and GPT 5.6 Sol.