At a Glance
| Vendor | Moonshot AI |
| Family | Kimi K series |
| Launched | Moonshot AI released Kimi K3 on July 16, 2026, succeeding Kimi K2.6. At 2.8 trillion total parameters it is the largest open-weight model released to date and the first into the 3-trillion-parameter class. Weights were scheduled for July 27, 2026 under a modified MIT license. |
| Context window | 1,000,000 tokens, with native image and video understanding alongside text. |
| Pricing | Hosted API pricing of $3 per million input tokens and $15 per million output tokens. Self-hosting is free under the modified MIT license, though the hardware required to serve 2.8T parameters puts that well beyond individual developers. |
| Access channels | Moonshot AI API, the Kimi consumer chat product, Hugging Face open-weight release, and inference providers adding K3 support through late July 2026. |
Notable Benchmarks
Kimi K3 scored 1,679 on Arena.ai's Frontend Code Arena, taking first place ahead of Claude Fable 5 at 1,631. Moonshot reports two architectural changes behind the gains: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). Sparse mixture-of-experts routing activates only a fraction of the 2.8T parameters per request.
Strengths
Frontier-competitive quality under an open licence, first-place Frontend Code Arena standing, 1M context with native multimodality, and MXFP4 quantisation support that makes serving more tractable than the raw parameter count suggests.
Limitations
Serving 2.8T parameters is a data-centre workload, not a workstation one, even with sparse activation and aggressive quantisation. Western enterprise governance concerns around Chinese-origin models persist regardless of licence terms, and the weights release date trailed the model announcement by 11 days.
Brand-Visibility Implications
Kimi K3 is the clearest evidence yet that the open-weight frontier has caught the closed frontier on specific capabilities. For brand visibility, the consequence is structural: every open-weight model that reaches frontier quality is a model your brand can be evaluated by inside an enterprise VPC, with no API call you can observe. The share of brand-relevant inference that is invisible to conventional monitoring grows with each release like this one. See Moonshot Kimi lineage and the local-LLM visibility blind spot.
How Presenc AI Tracks This Model
Presenc AI monitors brand visibility on Moonshot AI's Kimi K series as part of continuous multi-platform AI visibility tracking. We sample Moonshot Kimi K3 across representative prompt sets daily, compare against competitor performance on the same prompts, and flag material mention-rate changes so brand teams can respond quickly when AI representation shifts.