Moonshot Kimi K1.52,000,000 Token ContextHierarchical NVMe KV Cache100% Needle AccuracyRadix Prefix Deduplication

Moonshot Kimi K1.5 2M Ultra-Long Context Studio

Orchestrate Moonshot AI's breakthrough 2-Million token context engine. Master hierarchical KV-cache storage across GPU HBM, host DRAM, and NVMe SSDs, radix prefix caching for instant multi-turn retrieval, and long-horizon reinforcement learning planning.

Max Context Window

2,000,000 Tokens

~8 Million English Words

Cache Hit Speedup

94.2% Hit Rate

Prefix caching bypasses prefill

Needle Retrieval

100.0% NIAH

Zero loss across 2M tokens

Prefill Latency

-88% Reduction

Radix tree cache reuse

🏛️ Hierarchical Multi-Tier KV-Cache Architecture

Kimi K1.5 2M Engine

⚙️ Long-Context & Cache Tiering Config

Target Context Window

KV-Cache Tiering & Offload Engine

Long-Horizon RL Reasoning Mode

Serving Hardware Backend

💡 FinOps & Cache Efficiency

By offloading inactive token KV-states to PCIe 5.0 NVMe drives and reusing radix prefix trees, Kimi K1.5 serves 2M token context at 91% lower VRAM hardware cost compared to uncompressed attention.

Loading Kimi K1.5 configuration...