Kimi K2.5: Visual Agentic Intelligence

Kimi K2.5 is Moonshot AI's open-weight multimodal model, built on Kimi K2 (a mixture-of-experts model with 1.04 trillion total and 32 billion active parameters) by continued training on about 15 trillion mixed text and vision tokens. Most of the report is about training text and vision together. For anyone building agents, the important part is Agent Swarm.

Most agents call one tool at a time, so a task that needs 300 searches takes 300 steps, and one context window holds everything they return. Agent Swarm gives the model two extra tools. create_subagent defines a sub-agent with its own system prompt. assign_task sends it work. Several sub-agents run in parallel, each with its own context, and each returns a single message to the orchestrator.

The orchestrator learns when to do this through Parallel-Agent Reinforcement Learning (PARL). Only the orchestrator is trained; sub-agents are frozen copies of an earlier checkpoint whose outputs count as observations. That avoids the problem of deciding which agent deserves credit for a final answer. The reward is the task score plus two temporary bonuses: one for creating sub-agents, so training does not collapse into a single serial agent, and one for sub-agents finishing their assignments, so the orchestrator does not spawn agents that do nothing. Both bonuses decay to zero by the end of training.

Latency is measured in critical steps: the orchestrator's steps plus, for each batch of parallel sub-agents, the steps of the slowest one. Splitting work only helps if it shortens the longest branch.

On WideSearch, the swarm reached a given score 3 to 4.5 times faster than a single agent, and scored 79.0% item-F1 against 72.7%.

Customer research tasks are often wide: pull the latest filings for 40 suppliers, check 200 contracts for a clause, compare pricing across 30 competitors. A single agent does them one at a time and runs out of context partway through. Agent Swarm is a lab's worked answer: decompose, run in parallel with separate contexts, merge summaries.

Three things transfer even if you never use Kimi.

Separate contexts are context management. Each sub-agent sees only its share; the orchestrator sees only results. The paper frames this as sharding context rather than truncating it, and on BrowseComp it beat the same model using discard-all context management.

Measure latency as the longest branch. Critical steps make the trade-off explicit: eight sub-agents where one handles every item is no faster than one agent.

Guard against fake parallelism. The finish-rate bonus exists because the orchestrator learned to spawn agents that did nothing, which inflated a parallelism score without helping.

What the paper does not tell you is the bill. It reports time saved and accuracy gained, but no total tokens, tool calls or cost for the swarm. Each BrowseComp sub-agent may make up to 100 tool calls. Faster wall-clock time can mean much more total work. Before recommending a swarm, measure total tokens per completed task next to latency.

Read the 17.8-point BrowseComp gain against its baseline: the single agent had no context management and failed whenever its context overflowed. With discard-all context management it scored 74.9%, so the swarm's edge there is 3.5 points.