One-line summary
On October 24, 2026 (early morning Beijing time), Moonshot AI launched Kimi K3.1 Pro, a new open-source model with about 2.8 trillion total parameters (reported by CCTV Finance as "3-trillion-class"), 104B active parameters per token, and a 1-million-token context window, remaining the largest open-source model by parameter count. It scored 55 on the Artificial Analysis Intelligence Index (4th overall); API pricing is $2 per million input tokens and $10 per million output tokens.
Event Overview
- When: October 24, 2026, early morning Beijing time (reported by CCTV Finance).
- Who: Beijing Moonshot AI Technology Co., Ltd. (Moonshot AI).
- What: The company released its next-generation open-source model Kimi K3.1 Pro. CCTV Finance describes it as "3-trillion-class parameter scale"; independent evaluator Artificial Analysis cites 2.8T total parameters. The model supports a 1-million-token context window and keeps Moonshot AI as the publisher of the largest open-source model by parameters.
- Pricing and release cadence: Artificial Analysis records API pricing of $2 per million input tokens, $10 per million output tokens, and $0.20 for cache hits, each about one-third below the previous K3 model. Open weights are expected in November 2026 (not published at launch).
Key Information
| Item | Details |
|---|---|
| Publisher | Moonshot AI (Beijing) |
| Model | Kimi K3.1 Pro |
| Architecture | Mixture-of-Experts (MoE) |
| Total parameters | About 2.8 trillion (CCTV Finance: "3-trillion-class") |
| Active parameters | 104B per token |
| Context window | 1 million tokens |
| Open-source positioning | Largest open-source model by parameters (official claim) |
| Intelligence Index | 55 on Artificial Analysis Intelligence Index, 4th overall (11 points above K3) |
| Coding benchmark | 67% on Terminal-Bench 4.0 (AA measurement; Anthropic Sonnet 5.5: 64%) |
| API pricing | $2 / $10 per million input/output tokens; $0.20 cache hits (about one-third below K3) |
| Weight availability | November 2026 (announced open-source; weights to be released) |
Capability Highlights
- Coding: CCTV Finance reports Kimi K3.1 Pro uses professional design and development tools together with web search and 3D generation for game and web production; long-horizon coding was strengthened through large-scale reinforcement learning.
- 3D creation: From a text description, sketch, or reference image, the model infers object shape, proportions, and spatial relations, operates professional modeling software to produce usable 3D models, and supports online preview, editing, and download for 3D printing.
- Finance: From information retrieval and cross-document data verification to building financial models and profitability analysis, the model delivers end to end; CCTV Finance says it outperformed leading closed-source models in two international evaluations.
- Safety pruning: Artificial Analysis notes the open-weight release deliberately removed cyber-related training data from mid-training and reinforcement learning; Moonshot's internal xdev-bench-hard score (pass@3) dropped from 90.5% to 81.0%.
Background and Impact
- This is Moonshot AI's latest record in open-source parameter scale after K3, continuing the "large total parameters + sparse activation (MoE)" approach: about 104B active parameters support 2.8T total parameters to control per-inference cost.
- Artificial Analysis expects K3.1 Pro to become the highest-scoring open-weights model once weights land in November; currently available open-source models score below its 55.
- At about two-thirds of K3's price, it competes directly with DeepSeek V4.1 Flash (off-peak input ¥1, output ¥4 per million tokens) and MiniMax M3.1 Flash (input ¥1, output ¥4 per million tokens) in the open-source flagship segment.