# Chinese startup Moonshot AI launches Kimi K3, ranks 1st on the Frontend Code Arena *Published:* 2026-07-16 *Author:* Farzan Hussain ![](https://bestforandroid.com/wp-content/uploads/2026/07/Moonshot-AI-Kimi-K3-Open-Frontier-Intelligence.jpg) The Beijing-based startup released Kimi K3, a 2.8 trillion-parameter model that Moonshot calls the first open model in the 3 trillion-parameter class. It runs on a new architecture the company built in-house, comes with a 1-million-token context window, and is live now inside the Kimi app, the Kimi Work desktop tool, and Kimi’s coding assistant. Moonshot’s own benchmark data shows K3 trailing Anthropic’s Claude Fable 5 and OpenAI’s GPT 5.6 Sol overall, though it beats most other models it was tested against. > Big news: Kimi-K3 by [@Kimi\_Moonshot](https://x.com/Kimi_Moonshot?ref_src=twsrc%5Etfw) is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. > > This is a 17-place jump from Kimi-k2.6 (#18 -> #1). > > In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics,… [pic.twitter.com/Oa6teaQnWp](https://t.co/Oa6teaQnWp) > > — Arena.ai (@arena) [July 16, 2026](https://x.com/arena/status/2077824029126504525?ref_src=twsrc%5Etfw) What Kimi K3 is actually good at -------------------------------- Moonshot didn’t just publish a spec sheet. It ran K3 through a kernel optimization test against four other models, including Fable 5 and GPT 5.6 Sol, and reported that K3 performed competitively with Fable 5 while clearly beating Opus 4.8, GPT 5.6 Sol, and GPT 5.5 on that specific task. On a coding benchmark called Terminal Bench 2.1, K3 scored higher than Fable 5, Opus 4.8, and GPT 5.5, though it still landed just behind GPT 5.6 Sol. These aren’t sweeping wins. They’re specific, and Moonshot is transparent that the full picture still favors the two Western frontier models. > Introducing Kimi K3: Open Frontier Intelligence > > 🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal > 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts > 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional… [pic.twitter.com/eFHEbdxn3P](https://t.co/eFHEbdxn3P) > > — Kimi.ai (@Kimi\_Moonshot) [July 16, 2026](https://x.com/Kimi_Moonshot/status/2077830229968683203?ref_src=twsrc%5Etfw) The part most companies wouldn’t admit -------------------------------------- Moonshot’s [own release notes](https://www.kimi.com/blog/kimi-k3) list K3’s weaknesses in plain language. The model can make unexpected decisions on its own during long tasks if it hits ambiguous instructions, and Moonshot says there’s a noticeable gap in overall user experience compared with Fable 5 and GPT 5.6 Sol. I’ve read a lot of launch posts. Most of them bury the bad news, if they mention it at all. Seeing a company put its own shortcomings next to its wins in the same document is rare enough that it’s worth pointing out on its own. What this means for the price you’re paying ------------------------------------------- Here’s where it gets interesting for anyone building on top of these models instead of just chatting with them. Moonshot’s official API pricing runs 30 cents per million tokens for cached input, 3 dollars for fresh input, and 15 dollars for output. Full open weights, the kind developers can download and run themselves, are scheduled to ship by July 27. Kimi K3 isn’t the model that dethrones Fable 5 or GPT 5.6 Sol. It’s the model that makes both of them a little more expensive to justify.