Kimi (Moonshot AI)

Moonshot AI releases Kimi K3.1 Pro, an open-weight 3T-class model with 1M-token context and improved coding, 3D modeling, and financial reasoning capabilities.

Moonshot AI introduces Kimi K3, claimed as first open 3T-class model with frontier coding/reasoning performance, native multimodality, and 1M-token context.

PerceptionBench measures atomic-level visual perception in multimodal models by evaluating core perceptual failures rather than high-level task performance.

Kimi K2.6 adds long-horizon coding, agent swarms, and proactive agent capabilities to Moonshot's open-source model, with Claw Groups research preview.

Kimi Agent Swarm is a self-organizing multi-agent system supporting up to 100 concurrent sub-agents for parallelized research, synthesis, and analysis tasks.

WorldVQA benchmarks atomic vision-centric world knowledge in multimodal LLMs, measuring factual grounding in visual scenes rather than general reasoning.

Kimi K2.5 introduces Visual Agentic Intelligence with state-of-the-art visual coding and autonomous agent swarm capabilities, trained on 15T tokens.

Kimi Vendor Verifier is a tool for verifying vendor trust chains, with limited technical details provided.

Moonshot AI releases Kimi K2 Thinking, an open-source model emphasizing chain-of-thought reasoning for complex problem-solving tasks.

Kimi K2 is a Mixture-of-Experts model with 32B activated/1T total parameters, achieving state-of-the-art coding, math, and frontier knowledge performance.