Kimi K3 Now Available: The World's Largest Open-Weight Model on Multos
Moonshot AI's 2.8-trillion-parameter Kimi K3 is now live on Multos. A frontier-class open model with a 1M token context window, ranked #1 on Frontend Code Arena and #4 overall on Artificial Analysis Intelligence Index.
We're excited to announce that Kimi K3 — Moonshot AI's 2.8-trillion-parameter flagship model — is now available on Multos. Released on July 16, 2026, K3 is the world's first open 3T-class model and delivers frontier-level performance that rivals GPT-5.6 Sol and Claude Opus 4.8, particularly for coding and long-horizon agent tasks.
What Makes Kimi K3 Special
Kimi K3 is a Mixture-of-Experts model with 2.8 trillion total parameters (16 of 896 experts active per token), a 1-million-token context window, native multimodal input (text, images, video), and always-on chain-of-thought reasoning. It was purpose-built for sustained, multi-step engineering sessions — reading entire repositories, running terminal commands, interpreting failures, and iterating autonomously over extended periods.
Benchmark Highlights
K3 ranks #4 overall on Artificial Analysis Intelligence Index (score ~57, behind Claude Fable 5 and GPT-5.6 Sol Max). It holds the #1 position on LMArena's Frontend Code Arena in blind human evaluation, ahead of Claude Fable 5 and GPT-5.6 Sol. It leads on SWE Marathon (sustained multi-step coding) and Program Bench (building programs from spec), and scores 76.8% on SWE-bench Verified and 93.5% on GPQA graduate-level science reasoning.
Best For
Kimi K3 excels at long-horizon coding sessions where the model needs to sustain focus across multiple files and iterations. Frontend and UI code generation (independently ranked #1). Multi-step agent workflows involving tool use and terminal operations. Large codebase analysis using its full 1M token context window. Multimodal debugging — comparing rendered UI against design screenshots.
Pricing and Tier
Kimi K3 is available on the Pro plan at $3/$15 per million tokens (input/output). Moonshot's API achieves over 90% cache-hit rates on coding workloads, which reduces effective input cost significantly for iterative sessions. On Multos, select it from the model dropdown as 'Kimi K3 (Novita)' or let Auto Router assign it when your task demands frontier-level reasoning.
How It Compares
K3 sits in the same cost bracket as GPT-5.6 Terra on Multos. It outperforms Claude Opus 4.8 and GPT-5.5 on most benchmarks, trades blows with GPT-5.6 Sol (winning on sustained coding, losing on single-pass repository comprehension), and trails only Claude Fable 5 on overall capability. For teams that want near-frontier intelligence at Pro-tier pricing, K3 is a compelling option.
Ready to Get Started?
Try Kimi K3 now — select it from the model dropdown in your chat session or let Auto Router choose the best model for your task.
Start Building FreeRelated Articles
New Models: Gemini 3.6 Flash & Gemini 3.5 Flash-Lite Now Available
Two new Google Gemini models are now live on Multos — Gemini 3.6 Flash for Pro users and Gemini 3.5 Flash-Lite for Starter and above.
Read more →Auto Mode: How Smart Routing Stretches Your Token Budget
Auto mode picks the cheapest capable model per task. Manual selection gives you control with transparent cost multipliers.
Read more →GPT-5.6 Sol, Terra & Luna Are Now Live on Multos
OpenAI's most capable model family just dropped — all three variants are already available on Multos. Here's what they mean for your builds.
Read more →