Moonshot presents Kimi K3 as an open model aimed at frontier coding. Official materials describe 2.8T parameters, native vision, 1M context, Kimi Delta Attention, Attention Residuals, and a sparse MoE with 16 active experts out of 896.
Full weights are promised for July 27, 2026, so independent verification is still ahead.
Levels of confirmation
Architecture parameters and Kimi API availability are confirmed by Moonshot. Strong benchmark claims, including that K3 trails only Claude Fable 5 and GPT-5.6 Sol and performs especially well on Frontend Code Arena, should still be read as early signals rather than settled market facts.
Frontend has become an important arena for a reason. The model is not judged only by unit tests. It must combine code, layout, visual hierarchy, accessibility, performance, and human preference.
Native vision and screenshot iteration begin to matter as much as raw code generation.
How to test it
Moonshot already had a strong agentic-model line: Kimi K2 Thinking showed long tool chains and test-time scaling. K3 moves competition into visual software engineering.
It should be tested with a real frontend harness, not a vague “make a website” prompt: same prompt, same screenshots, browser checks, human preference, and comparison with a closed model.
New Runtime Read
If Kimi K3 consistently wins where the result is visible to the eye and checked in a browser, that matters more than ordinary benchmark marketing.
In that case, open model competition arrives where developers feel quality immediately: in the interface, not only in a synthetic score.