One-sentence summary
Kimi K3 and K3.1 Pro are both 2.8-trillion-parameter MoE flagships from Moonshot AI: K3's weights opened in July 2026, while K3.1 Pro (released October 24, 2026) is a post-training upgrade with weights planned before November 4, 2026 and a Highspeed variant running 2-3x faster.
Model Introductions
Kimi K3 is the previous-generation flagship released by Moonshot AI on July 27, 2026. With 2.8 trillion parameters built on Kimi Delta Attention (KDA) and Attention Residuals, it is the world's first open 3T-class model. Moonshot AI says Kimi models held the open-model size ceiling in 9 of the past 12 months (July 2025 to July 2026).
Kimi K3.1 Pro, released October 24, 2026, is the first model of the K3 series, a post-training upgrade of K3 focused on long-horizon coding, 3D modeling and spatial reasoning, finance research, and long-horizon agent execution. A Highspeed variant was announced the same day.
Core Parameter Comparison
| Dimension | Kimi K3 | Kimi K3.1 Pro |
|---|---|---|
| Release | 2026-07-27 (K3 Open Day) | 2026-10-24 |
| Total parameters | 2.8 trillion | 2.8 trillion (vendor: "3-trillion-class") |
| Active parameters per token | 104 billion | 104 billion |
| Architecture | KDA + Attention Residuals; Stable LatentMoE (16 of 896 experts active) | Inherits K3 architecture; post-training with RL, long-context reasoning, task scheduling |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Input modalities | Text, image, video | Text, image, video |
| Weight status | Open (2026-07-27) | Planned before 2026-11-04 |
| Variants | None | K3.1 Pro Highspeed (~2-3x output speed, 2x price) |
| Coding benchmarks | FrontierSWE 81.2; ProgramBench 77.8; SWE Marathon 42.0; Terminal Bench 2.1 88.3 | Terminal-Bench 4.0 65.7; Kimi Code Bench 3.0 Long-horizon 61.2 (internal) |
| 3D / finance | General-purpose focus | BenchCAD (w/ tools) 92.7; OfficeQA Pro v2 64.4 (vendor claims SoTA); Frontier Finance 58.6 |
| API access | Recharge-gated on open platform (min CNY 10) | Kimi web, App, Kimi Work, Kimi Code, Kimi API |
| Fact error rate (internal stress test) | 23.4% (Kimi Hallu Bench Hard single-fact) | 8.3% |
Key Differences
Positioning and Training Route
K3 emphasized "open frontier intelligence," open-sourcing a 3T-class model and publishing KDA, Attention Residuals, and Stable LatentMoE details with its tech report. K3.1 Pro's upgrade is not about scale (same 2.8T) but about converting larger reinforcement-learning investment into sustained execution on complex tasks. Moonshot AI reports internal financial modeling up ~14%, finance research up ~20%, and business operations around 4.2x vs K3.
Benchmark Results
K3's public coding numbers include FrontierSWE 81.2, ProgramBench 77.8, Terminal Bench 2.1 88.3, and SWE Marathon 42.0 (kimi.ai tech blog, updated October 9, 2026). K3.1 Pro scored 65.7% on Terminal-Bench 4.0, above GPT-6 Astra's 58.2% and below Claude Opus 5.5's 66.4%; its internal Kimi Code Bench 3.0 Long-horizon is 61.2%, above K3's 55.95%. The two generations use different benchmark versions (Terminal-Bench 4.0 vs 2.1, Kimi Code Bench 3.0 vs 2.0), so cross-generation numbers are not directly equivalent.
Variants and Commercialization
K3.1 Pro adds the Highspeed variant (~2-3x output speed, 2x price, comparable intelligence) and the Kimi Hosted Agents managed-agent service. No equivalent high-speed variant was publicly reported for K3.
Open-Source Rhythm and Security Assessment
K3 weights are fully released. K3.1 Pro follows a "services first, weights later" plan, with full weights planned before November 4, 2026. Before opening, Moonshot AI disclosed it removed cyber attack-defense-related data from mid-training and RL stages; xdev-bench-hard dropped from 90.5% (before removal) to 81.0%.
Buying Guidance
Teams needing local deployment, fine-tuning, or weight-level research can use the already-open Kimi K3. API use cases needing stronger long-horizon coding, 3D, and finance can use K3.1 Pro (via API until weights open). For latency-sensitive production, evaluate K3.1 Pro Highspeed first. Both generations share the same context window and active parameters, so migration cost is mostly prompt adaptation and benchmark-caliber alignment.
Data Sources and References
- 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
- Kimi official (K3 Open Day announcement): https://www.kimi.com/news/kimi-k3-open-source
- Wall Street CN (2026-10-24): http://m.toutiao.com/group/7700018343629046308/
- AIToolsRecap (2026-10-24, citing Artificial Analysis): https://aitoolsrecap.com/Blog/ai-news-october-24-2026