Moonshot drops a 1T-parameter MoE that sees and swarms

What it does Kimi K2.5 is a 1-trillion-parameter Mixture-of-Experts model (32 billion active) trained continually on roughly 15 trillion mixed visual and text tokens. It reads images and video, writes code from UI mockups or workflow diagrams, and can flip between instant replies and extended thinking. The repo appears to be a model-card release: it publishes architecture details, benchmark tables, and a research summary rather than inference code or weights.

The interesting bit Instead of scaling a single monolithic agent, Kimi K2.5 spawns a self-directed “agent swarm”—dynamically instantiating domain-specific sub-agents to run parallel sub-tasks. That turns one big model into a temporary team of specialists, which is a neat answer to the “one model can’t do everything” problem.

Key highlights

• Sparse MoE architecture: 1T total parameters, only 32B activated per token, 384 experts with 8 selected per forward pass.
• Native multimodal pretraining on vision-language tokens; includes a 400M-parameter MoonViT vision encoder.
• 256K context window using MLA (Multi-head Latent Attention).
• “Coding with Vision”: generates code directly from visual specifications like UI designs and video workflows.
• Licensed under Modified MIT.

Caveats

• The repository appears to contain documentation and benchmarks only; no model weights, training code, or inference stack are visible in the provided README.
• Benchmark tables compare against models with unfamiliar version labels (e.g., “GPT-5.2”, “Claude 4.5 Opus”); the README does not clarify whether these are public releases or internal codenames.

Verdict Worth a look if you are researching multimodal agent architectures or MoE scaling laws. Skip it if you need downloadable weights or a local inference server today.