Moonshot AI

<p>Authored by the Kimi Team at <strong>Moonshot AI</strong> (including <a href="/labs/moonshot-ai">Su Jianlin</a>), this paper introduces a fundamen…

Open-source multimodal agentic model designed to advance general agentic intelligence. Features joint text-vision optimization and Agent Swarm, a self…

LLM abstraction layer for building agent applications.

Hybrid linear attention architecture (KDA + MLA). 3B active / 48B total MoE model with 75% KV-cache reduction and 6x throughput at 1M context.

Middleware tool for updating model weights dynamically in inference engines.

Landmark 1-trillion parameter MoE model family (32B active) released as open-weight. 384 experts with 8 activated per token. Pre-trained on 15.5T toke…

Benchmarks developed to evaluate an agent's ability to browse the web and synthesize information from multiple sources.

Multi-language SDK (Python, Go, Node.js) for programmatically interacting with Kimi's agentic features.

Command-line agent designed for autonomous coding and system tasks.

Open-source SWE LLM achieving 60.4% on SWE-bench Verified. Uses agentless training as skill prior for SWE-Agents.

Specialized iteration for autonomous research.

Specialized library written in Go for high-performance AI infrastructure.

Universal audio foundation model for understanding, generation, and conversation. Trained on 13M+ hours of audio data.

Formal reasoning model for Lean 4 theorem proving trained via RL from Qwen2.5-72B. Achieves 80.7% on miniF2F with pass@8192.

Cleaned and corrected version of the popular formal reasoning benchmark, released alongside Kimina-Prover.

Open-source vision-language model with Moonlight MoE LLM (2.8B active / 16B total) and 400M MoonViT encoder. Kimi-VL-Thinking achieves 64.0 on MMMU.

Optimizer based on matrix orthogonalization. Research demonstrating scalability for large language model training.

Lightweight "Pareto-frontier" models trained using the Muon optimizer. Available in 3B and 16B variants.

Mixture of Block Attention mechanism for efficient long-context processing.

Reasoning-focused model using reinforcement learning, claimed to match OpenAI's o1-preview in math and coding. Technical report details scaling RL wit…

A KVCache-centric disaggregated architecture for LLM serving. Won Best Paper at FAST 2025.

Update to the proprietary model supporting a 2-million-character context window.

First model to support 128,000 to 200,000 tokens of lossless context, a world-first at the time. Included moonshot-v1-8k, 32k, and 128k API variants p…