One-sentence summary
This page aggregates structured data for Moonshot AI's flagship Kimi K3.1 Pro, released October 24, 2026: architecture, context window, input modalities, public benchmark scores, API pricing, and the open-weight plan, all as of October 24, 2026.
Data Description
Sources: Moonshot AI official releases (Kimi website, Kimi API platform documentation, Kimi K3 tech blog), a Wall Street Journal-style report by Wall Street CN dated October 24, 2026, and AIToolsRecap's October 24, 2026 summary citing Artificial Analysis. Definitions: total parameters are reported as 2.8 trillion (the vendor's phrasing is "3-trillion-class"); active parameters are 104 billion per token; prices are per million tokens, in CNY per Wall Street CN and in USD per AIToolsRecap citing Artificial Analysis; benchmark scores are official or reported raw values without conversion.
Core Data
| field | field meaning | value / notes |
|---|---|---|
| model_name | Model name | Kimi K3.1 Pro (variant: Kimi K3.1 Pro Highspeed) |
| release_date | Release date | 2026-10-24 |
| total_params | Total parameters | 2.8 trillion (vendor: "3-trillion-class") |
| active_params | Active parameters per token | 104 billion (~104B) |
| architecture | Architecture | Mixture of Experts (MoE) on Kimi Delta Attention (KDA) and Attention Residuals |
| context_length | Context window | 1,000,000 tokens |
| input_modalities | Input modalities | Text, image, video |
| output_modalities | Output modalities | Text |
| weight_open_date | Open-weight plan | Before 2026-11-04 |
| access_channels | Access channels | Kimi web, App, Kimi Work, Kimi Code, Kimi API |
| benchmark | metric note | score |
|---|---|---|
| Terminal-Bench 4.0 | Agentic terminal programming | 65.7% |
| Kimi Code Bench 3.0 Long-horizon | Internal benchmark, long-horizon engineering | 61.2% (K3 was 55.95%) |
| BenchCAD (w/ tools) | CAD code 3D reconstruction | 92.7% |
| OfficeQA Pro v2 | Financial document analysis | 64.4 (vendor claims SoTA) |
| Frontier Finance | Finance research | 58.6 |
| Agents' Last Exam | Agent capability | 43.2 |
| Humanity's Last Exam (Full w/ tools) | Broad knowledge | 79.0 |
| Terminal-Bench Science 0.1 | Scientific terminal tasks | 68.1 |
| GDPVal-AA v2.1 | General agent Elo | 1784.9 |
| OSWorld V2.1 | Computer-use agent | 81.8 |
| AA Intelligence Index | Artificial Analysis intelligence index | 55 (4th among public models) |
| price_item | billing item (per million tokens) | CNY | USD |
|---|---|---|---|
| input_miss | Input (cache miss) | 13 CNY | $2 |
| input_hit | Input (cache hit) | 1.3 CNY | $0.20 |
| output | Output | 68 CNY | $10 |
| highspeed_multiplier | Highspeed variant price | 2x standard | 2x standard |
| fact_item | factuality / safety metric | value / notes |
|---|---|---|
| hallucination_rate | Kimi Hallu Bench Hard single-fact error rate | 8.3% (K3: 23.4%) |
| xdev_hard_after | xdev-bench-hard (after cyber data removal) | 81.0% |
| xdev_hard_before | xdev-bench-hard (before removal) | 90.5% |
| internal_fin_model | Internal financial modeling gain vs K3 | ~14% |
| internal_fin_research | Internal finance research gain vs K3 | ~20% |
Field Notes
model_name/weight_open_date: the weight release is planned and subject to final license terms. architecture: KDA and Attention Residuals are Moonshot AI's in-house components; KDA is a hybrid linear attention mechanism. Benchmark caliber: Kimi Code Bench series is an internal benchmark; on Terminal-Bench 4.0, Moonshot reports 65.7% for K3.1 Pro, 58.2% for GPT-6 Astra, and 66.4% for Claude Opus 5.5; on BenchCAD w/ tools, Claude Opus 5.5 scores 96.2% and GPT-6 Astra 95.9%. price_item: USD figures come from AIToolsRecap citing Artificial Analysis ($2 input, $10 output, $0.20 cache hit); CNY figures come from Wall Street CN (13 CNY input miss, 1.3 CNY input hit, 68 CNY output); the two sets describe the same pricing scheme. hallucination_rate is Moonshot AI's internal stress-test data, which the company notes uses deliberately hard scenarios and does not represent overall daily error rates.
Data Sources and References
- Wall Street CN (2026-10-24): http://m.toutiao.com/group/7700018343629046308/
- Kimi official (K3 tech blog): https://www.kimi.ai/blog/kimi-k3
- Kimi API platform (K3 docs): https://platform.kimi.com/docs/guide/kimi-k3-quickstart
- AIToolsRecap (2026-10-24, citing Artificial Analysis): https://aitoolsrecap.com/Blog/ai-news-october-24-2026
- CCTV Finance (2026-10-24): http://m.toutiao.com/group/7700021695326847522/