Moonshot Kimi K3.1 Pro (3T MoE)
The largest open-weight Mixture-of-Experts foundation model ever released: 3 trillion total parameters with 96B active per token, trained via the MuonClip optimizer with Multi-Head Latent Attention (MLA).
Interactive Token Cost Estimator (Kimi K3.1 Pro (3T))
GPU Hardware Compatibility Matrix (Kimi K3.1 Pro (3T) at 16K Context)
| GPU / Hardware Profile | VRAM | Q4_K_M (4.8-bit) | Q8_0 (8-bit) | Est. Speed (Q4) |
|---|---|---|---|---|
| RTX 5060 Ti 16GB 448 GB/s | 16 GB | REQUIRES MULTI-GPU (1802.91 GB) | REQUIRES MULTI-GPU (3182.99 GB) | ~1 tok/s |
| RTX 4090 24GB 1008 GB/s | 24 GB | REQUIRES MULTI-GPU (1802.91 GB) | REQUIRES MULTI-GPU (3182.99 GB) | ~2 tok/s |
| RTX 5090 32GB 1792 GB/s | 32 GB | REQUIRES MULTI-GPU (1802.91 GB) | REQUIRES MULTI-GPU (3182.99 GB) | ~4 tok/s |
| M4 Max 64GB 546 GB/s | 52 GB | REQUIRES MULTI-GPU (1802.91 GB) | REQUIRES MULTI-GPU (3182.99 GB) | ~1 tok/s |
| 2× 24GB (48GB) 920 GB/s | 48 GB | REQUIRES MULTI-GPU (1802.91 GB) | REQUIRES MULTI-GPU (3182.99 GB) | ~2 tok/s |
| H100 80GB 3350 GB/s | 80 GB | REQUIRES MULTI-GPU (1802.91 GB) | REQUIRES MULTI-GPU (3182.99 GB) | ~8 tok/s |
| M4 Ultra 192GB 819 GB/s | 164 GB | REQUIRES MULTI-GPU (1802.91 GB) | REQUIRES MULTI-GPU (3182.99 GB) | ~2 tok/s |
| 8× H200 (1,128GB) 4800 GB/s | 1128 GB | REQUIRES MULTI-GPU (1802.91 GB) | REQUIRES MULTI-GPU (3182.99 GB) | ~12 tok/s |
Architectural Strengths & Target Workloads
- Open-weight model rivaling proprietary frontier systems on SWE-bench, AIME, and long-horizon agent benchmarks
- Multi-Head Latent Attention (MLA) compresses KV-cache by 84% across 2M context
- 96B active parameters deliver fast per-token generation once distributed across NVLink H200/B200 nodes
Primary Engineering Workloads: Sovereign / On-Prem Frontier Cluster · Long-Horizon Agent Swarms · Deep Codebase Synthesis.
Frequently Asked Questions (Kimi K3.1 Pro (3T))
How much does Moonshot Kimi K3.1 Pro (3T MoE) cost per 1M tokens?
Moonshot Kimi K3.1 Pro (3T MoE) is priced at $0.90 per 1M uncached input tokens, $0.12 per 1M cached input tokens, and $3.20 per 1M output tokens (Moonshot API / Together / Fireworks or Self-Hosted Multi-Node vLLM).
What is the context window of Moonshot Kimi K3.1 Pro (3T MoE)?
It supports up to 2,000,000 tokens (2,000,000 tokens) and a maximum generation output of 131,072 tokens.
Moonshot Releases Kimi K3.1 Pro: 3-Trillion-Parameter Model Beats GPT-6 Astra on Terminal-Bench 4.0
With a 1-million-token context window and open weights scheduled for November 4, Kimi K3.1 Pro closes the gap between open-weight models and Claude Opus 5.5.
Read Full Technical Dispatch (6 min read) →