Your sovereign, private Kimi K3 Priority
Kimi K3 is Moonshot AI's flagship open-weight, natively multimodal agentic model: a 2.8T-parameter MoE (104B active) built on Kimi Delta Attention and Attention Residuals, with native vision, a 1M-token context window and interleaved thinking with configurable reasoning effort. It is designed for frontier intelligence across long-horizon coding, deep research and agentic knowledge work, sustaining multi-hour engineering sessions and orchestrating terminal tools with minimal oversight. Served on the priority service tier. A reasoning model in the ultra tier, served on ToothFairyAI's AU, EU and US endpoints — private by default, billed in Units of Intelligence at the published rates below. No training on your data, no vendor lock-in.
How Kimi K3 Priority compares
Independent intelligence index for Kimi K3 Priority against every peer in the Ultra tier (and the tier above where needed) — same source, same tests.
Throughput — tokens per second, median
Latency — end-to-end seconds for a 500-token answer
Capability indexes
Domain-weighted agentic performance for Kimi K3 Priority.
Capability index vs ultra peers
The evaluations behind the index
Independent benchmark results for Kimi K3 Priority — Elo scores as published; pass rates shown as percentages.
Per-evaluation economics
| Evaluation | Score | Time / task | Output tokens / task |
|---|---|---|---|
| AA-Briefcase | 1504.04 | 49.5 min | 119873 |
| GDPval-AA | 1523.97 | 24.5 min | 59382 |
| AutomationBench-AA | 58.3% | 8.2 min | 19920 |
| Terminal-Bench 4.0 | 12.6% | 50.8 min | 122899 |
| SciCode | 59.5% | 1.5 min | 3520 |
| Humanity's Last Exam | 46.9% | 10.8 min | 26162 |
| GDP.pdf | 22.0% | 4.0 min | 9597 |
| CritPt | 23.4% | 24.5 min | 59238 |
| AA-Omniscience | 19.70 | 3.6 min | 8813 |
| AA-LCR (long context) | 88.7% | 0.6 min | 1529 |
Run Kimi K3 Priority privately
One key, one endpoint, three jurisdictions.
- Size tierUltra
- Input rate6.30 UoI/1M
- Output rate31.50 UoI/1M
- Cached input rate0.63 UoI/1M
- Context window800K tokens
- Max output120K tokens
- ReasoningYes — emits thinking traces
- Vision & video inputSupports image input
- Tool callingSupported
- Data residencyAU, EU and US endpoints — processing stays inside the region you call