Agent Swarm Cuts Complex Task Latency Up To 4.5×

Multimodal models can recognize images or reason over text, but useful agents must combine those abilities across long, tool-mediated tasks without letting a single context window become the bottleneck. Kimi K2.5 targets that operational problem rather than treating vision as a separate add-on: the authors jointly train text and vision, then use a parallel orchestration layer, Agent Swarm, to split complex work into bounded subproblems.

The paper reports broad results in reasoning, perception, computer use, video understanding, and agentic search. Its strongest evidence is not a single aggregate score but the combination of head-to-head benchmark tables and controlled comparisons between a swarm and a single-agent configuration. Still, most of the reported comparisons are against proprietary systems or the authors' earlier Kimi models, so the practical value is clearer than the source of every gain.

Core Contribution

Kimi K2.5 combines a multimodal post-training recipe with Agent Swarm. The training recipe comprises joint text-vision pre-training, zero-vision supervised fine-tuning, and joint text-vision reinforcement learning. The stated purpose of zero-vision SFT is to preserve text-only capability while the model is trained for visual work; the later joint RL stage is intended to make cross-modal reasoning useful in agent trajectories rather than only in static question answering.

Agent Swarm is the more distinct systems contribution. Instead of assigning a long task to one agent with one growing context, an orchestrator creates heterogeneous subagents as task structure emerges. Each receives a semantically isolated, local context and returns selected outputs to the orchestrator. That differs from reactive context compression methods, which discard or summarize accumulated history after overflow: the paper frames decomposition and selective routing as proactive context management.

Technical Approach

The orchestrator learns adaptive policies for creating and scheduling subagents rather than following a fixed task tree. In the paper's examples, this permits parallel source retrieval, independent verification, and extraction work, followed by aggregation by a main agent. The claimed benefit is twofold: independent workers avoid contaminating the central context, and concurrent execution prevents completion time from rising with each sequential tool call.

The visual-agent examples make the intended division of labor concrete. For a 24-hour, 32-video Black Myth: Wukong playthrough at 1080p, the system dispatches 32 subagents to analyze clips and produces a chronological HTML showcase. Other examples use tools for binary image segmentation and breadth-first search in a maze, pixel-level color segmentation for a pie chart, and image differencing for spot-the-difference puzzles. These are qualitative demonstrations of tool use and decomposition, not controlled accuracy measurements.

Results and Analysis

On reasoning benchmarks, Table 5 reports Kimi K2.5 at 96.1 on AIME 2025, 94.7 on HMMT February 2025, 91.1 on HMMT November 2025, and 81.8 on GPQA Diamond. Against the listed Kimi K2 Thinking baseline, those are gains of 1.2, 5.3, 1.9, and 2.3 points respectively; it also uses fewer average output tokens on three of those four tests. The comparison is uneven, however: Gemini-3.0 Pro exceeds K2.5 on several rows, including HLE-Text, where K2.5 scores 31.5 versus 38.4.

The multimodal evaluation is similarly broad. The paper reports 78.5% on MMMU-Pro, 71.2% on SimpleVQA, 46.5% on WorldVQA, 84.2% on MathVista, 90.1% on MathVista mini, 77.5% on CharXiv, 92.3% on OCRBench, and 88.8% on OmniDocBench 1.5. For longer video, it reports 75.9% on LVBench and 79.8% on LongVideoBench after feeding more than 2,000 frames. These figures support competence across many modalities, but the paper does not establish which part of the training recipe causes each improvement.

The Agent Swarm comparison is more diagnostic. On BrowseComp, swarm reaches 78.4 versus 60.6 for single-agent Kimi K2.5, a 17.8-point gain; WideSearch reaches 79.0 versus 72.7, and the in-house Swarm Bench reaches 58.3 versus 41.6. On WideSearch, the authors report $3\times$ to $4.5\times$ lower execution time at target Item-F1 levels from 30% to 70%. That is a meaningful operational result if parallel tool capacity is available, though it does not measure the added infrastructure cost or behavior under resource contention.

Caveats in Practice

The evidence favors the claim that orchestration helps on search-heavy, decomposable tasks. It is less conclusive for general agency: one benchmark is internally developed, many task results lack error bars or cost accounting, and qualitative visual demonstrations do not substitute for held-out quantitative tests. The model checkpoint is described as released, but the visible paper material does not provide a verifiable repository or download URL.