The Breakthrough: 100 Sub-Agents, 1,500 Tool Calls, 4.5x Speed Increase—Native Multimodal AI Redefining Agentic Intelligence

Kimi K2.5 represents the most powerful open-source model to date, achieving state-of-the-art coding and vision capabilities through native multimodal architecture trained on approximately 15 trillion mixed visual and text tokens. The Agent Swarm Innovation: Self-directed orchestration of up to 100 sub-agents executing parallel workflows across up to 1,500 tool calls, reducing execution time by 4.5x compared to single-agent setups—automatically created and coordinated without predefined workflows. The Coding Revolution: Strongest open-source coding model with exceptional front-end capabilities, turning simple conversations into complete interactive layouts with rich scroll-triggered animations, excelling at visual debugging by reasoning over images/video to improve image-to-code generation and lower barriers for visual intent expression. The Cost Advantage: Delivering strong performance on agentic benchmarks (HLE, BrowseComp, SWE-Verified) at fraction of competitor costs. The Office Productivity: Handling high-density large-scale work end-to-end, reasoning over massive inputs, coordinating multi-step tool use, delivering expert-level documents/spreadsheets/PDFs/slides through conversation—59.3% improvement on AI Office Benchmark and 24.3% improvement on General Agent Benchmark over K2 Thinking. The PARL Training: Parallel-Agent Reinforcement Learning with trainable orchestrator decomposing tasks into parallelizable subtasks, frozen subagents executing concurrently, staged reward shaping preventing serial collapse, critical-steps metric forcing parallel strategies. Availability: Via Kimi.com, Kimi App, API (platform.moonshot.ai), and Kimi Code—four modes (K2.5 Instant, Thinking, Agent, Agent Swarm Beta).

Part I: The Multimodal Foundation

Massive-Scale Vision-Text Joint Pretraining

Training Corpus: Approximately 15 trillion mixed visual and text tokens

Architecture: Native multimodal model from ground up

Key Insight: “At scale, the trade-off between vision and text capabilities disappears—they improve in unison”

Result: State-of-the-art performance in both coding and vision tasks

Paradigm Shift: Vision and text not competing but complementary

The Unified Capability Emergence

Traditional Approach: Separate text and vision models

Kimi K2.5 Innovation: Single model excelling at both

Synergistic Learning: Visual reasoning enhancing code understanding

Practical Impact: Seamless multimodal workflows

Part II: Coding with Vision—The Front-End Revolution

Conversational Interface Generation

Capability: Turning simple conversations into complete front-end interfaces

Features Implemented:

• Interactive layouts
• Rich animations
• Scroll-triggered effects
• Complex UI components

Single Prompt Power: Complete implementations from minimal descriptions

Example: Image-gen tool integration producing fully functional interfaces

Developer Impact: Dramatically reduced front-end development time

Visual Debugging Breakthrough

The Innovation: Reasoning over images and video for code generation

Image-to-Code: Converting visual designs directly to implementation

Video-to-Code: Reconstructing websites from video demonstrations

Example Workflow:

1. Record video of desired website behavior
2. Feed to K2.5
3. Receive complete code reconstruction
4. Iterate based on visual feedback

Barrier Reduction: Users expressing intent visually instead of technical specifications

Autonomous Visual Iteration

Kimi Code Integration: Terminal-based tool integrating with VSCode, Cursor, Zed

Open Source: Freely available codebase

Multimodal Input: Supports images and videos

Auto-Discovery: Automatically finds and migrates existing skills and MCPs

Example – Matisse's La Danse Translation:

1. Visual input: Famous painting aesthetic
2. Documentation lookup: Kimi App design guidelines
3. Visual inspection: K2.5 checking own output
4. Autonomous iteration: Refining until aesthetically matching
5. End-to-end result: Art-inspired webpage created autonomously

The Breakthrough: AI visually debugging its own work without human intervention

Real-World Software Engineering

Kimi Code Bench: Internal benchmark covering diverse end-to-end tasks

Task Categories:

• Building from scratch
• Debugging existing code
• Refactoring for improvements
• Testing implementation
• Scripting automation

Language Coverage: Multiple programming languages

K2.5 vs K2 Improvement: Consistent and meaningful gains across all task types

Production Readiness: Strong performance on real-world engineering workflows

Visual Reasoning Example

Puzzle Solving: K2.5 analyzing visual puzzle

Code-Based Marking: Using code to mark shortest path solution

Integration: Vision understanding + code generation + logical reasoning

Practical Applications:

• Algorithm visualization
• Game development
• Educational tools
• Interactive problem solving

Part III: Agent Swarm—Scaling Out, Not Just Up

The Paradigm Shift

Traditional Scaling: Single agent with more compute (scaling up)

K2.5 Innovation: Multiple coordinated agents (scaling out)

Research Preview: Agent Swarm currently in beta

Shift Significance: From sequential to parallel agentic execution

The Technical Architecture

Orchestrator Agent: Trainable coordinator (not frozen)

Sub-Agents: Up to 100 dynamically instantiated (frozen during execution)

Task Decomposition: Breaking complex tasks into parallelizable subtasks

Dynamic Instantiation: Sub-agents created on-demand for specific roles

Example Roles:

• AI Researcher
• Physics Researcher
• Fact Checker
• Data Analyst
• Code Reviewer

No Predefined Workflows: Entirely self-directed coordination

Parallel-Agent Reinforcement Learning (PARL)

The Challenge: Training reliable parallel orchestrator

Problem 1 – Delayed Feedback: Sparse rewards from independently running sub-agents

Problem 2 – Non-Stationary: Sub-agent behaviors changing during training

Problem 3 – Serial Collapse: Orchestrator defaulting to single-agent despite parallel capacity

The Solution – Staged Reward Shaping:

Reward Function:

R_t = λ_aux(e) · r_parallel + (1 - λ_aux(e)) · (I[success] · Q(τ))
      ↑                        ↑
  instantiation reward    task-level outcome

Annealing Schedule: λ_aux decreases from 0.1 → 0.0 over training

Early Training: Auxiliary reward r_parallel incentivizes sub-agent instantiation

Late Training: Focus shifts to end-to-end task quality Q(τ)

Prevents: Degenerate solutions where parallelism exists nominally but not effectively

The Critical Steps Metric

Traditional Metric: Total steps counted

Problem: Doesn't capture parallel execution benefits

Critical Steps Definition:

CriticalSteps = Σ(S_main(t) + max_i S_sub,i(t))

Components:

• S_main(t): Orchestration overhead at time t
• max_i S_sub,i(t): Slowest sub-agent at time t

Inspiration: Critical path in parallel computation theory

Forcing Function: Spawning more subtasks only helps if shortening critical path

Result: Genuine parallel strategies emerge during training

Performance Improvements

End-to-End Runtime Reduction: Up to 80%

Speedup Factor: 3x–4.5x compared to single-agent execution

Critical Steps Reduction: 3x–4.5x fewer steps to achieve target performance

Scaling Behavior: Savings increase as task complexity rises

Wall-Clock Impact: 4.5x time reduction via parallelization

Complex Workloads: Enables longer-horizon tasks previously impractical

Execution Capacity

Maximum Sub-Agents: 100 concurrent

Maximum Tool Calls: 1,500 coordinated steps

Coordination Complexity: Automatic orchestration without manual workflow design

Benchmark Performance: Strong results on HLE, BrowseComp, SWE-Verified

Cost Efficiency: Fraction of competitor costs while maintaining performance

Training Progress Visualization

Smooth Reward Increase: Gradual improvement throughout training

Parallelism Level: Gradually increasing agent coordination

Convergence: Stable final performance without collapse

Reliability: Production-ready coordination mechanisms

Part IV: Office Productivity Revolution

Real-World Knowledge Work

Target: High-density, large-scale office tasks

End-to-End Handling: From input to finished deliverable

Output Formats:

• Microsoft Word documents
• Excel spreadsheets
• PDF files
• PowerPoint slide decks

Interface: All through natural conversation

Advanced Office Capabilities

Word Processing:

• Adding annotations
• Complex formatting
• Long-form content (10,000+ words)

Spreadsheet Mastery:

• Financial model construction
• Pivot Table creation
• Advanced formulas

PDF Generation:

• LaTeX equation writing
• Professional formatting
• 100+ page documents

Presentation Creation:

• Slide deck generation
• Visual design
• Content organization

Internal Expert Productivity Benchmarks

AI Office Benchmark: Evaluates end-to-end Office output quality

General Agent Benchmark: Measures multi-step production workflows against human experts

K2.5 vs K2 Thinking Improvements:

• 59.3% improvement on AI Office Benchmark
• 24.3% improvement on General Agent Benchmark

Real-World Focus: Tasks professionals actually perform daily

Expert-Level Output: Matching or exceeding human professional quality

Time Compression

Previous Reality: Tasks taking hours or days

K2.5 Performance: Minutes to completion

Productivity Multiplier: 10x-100x time savings potential

Workflow Integration: Seamlessly fitting into existing processes

Professional Impact: Redefining knowledge worker productivity

Part V: Benchmark Performance Deep Dive

Coding Benchmarks

SWE-Bench Series (Verified, Multilingual, Pro):

• Minimal toolset (bash, createfile, insert, view, strreplace, submit)
• Tailored system prompts
• Non-thinking mode optimal
• Averaged over 5 independent runs

Terminal-Bench 2.0:

• Default Terminus-2 agent framework
• JSON parser provided
• Non-thinking mode for compatibility

CyberGym: Claude Opus 4.5 comparison under non-thinking setting

Kimi Code Bench: Strong improvements across all task categories

Vision Benchmarks

MMMU-Pro: Official protocol, input order preserved, images prepended

WorldVQA: Atomic vision-centric world knowledge evaluation (github.com/MoonshotAI/WorldVQA)

OmniDocBench: Score = (1 – normalized Levenshtein distance) × 100

ZeroBench (with tools): Multi-step reasoning with 24k tokens per step, 30 max steps

Averaging: 3 runs (avg@3) for consistency

Agentic Search Benchmarks

Tools Equipped: Search, code-interpreter, web-browsing

Context Management: No management except BrowseComp (discard-all strategy)

Context Overflow: Tasks exceeding limit counted as failed

System Prompts: Emphasizing deep and proactive tool use

Averaging: 4 runs (avg@4) for Seal-0 and WideSearch

Reasoning Benchmarks

HLE (Text & Image):

• Full set: Text 31.5, Image 21.3 (without tools)
• Full set: Text 51.8, Image 39.8 (with tools)
• 96k token completion budget
• Hugging Face access blocked (prevent data leakage)

AIME 2025: 96k token budget, avg@32 (32 runs)

HMMT 2025 (Feb): 96k token budget, avg@32

GPQA-Diamond: 96k token budget, avg@8

IMO-AnswerBench: 96k token budget

Long-Context Performance

AA-LCR: Averaged over 3 runs (avg@3)

LongBench-V2: Identical prompts, input standardized to ~128k tokens

Context Length: 256k tokens supported

Part VI: Access and Availability

Four Modes Available

K2.5 Instant: Fast responses for simple queries

K2.5 Thinking: Extended reasoning for complex problems

K2.5 Agent: Tool-augmented execution with preconfigured capabilities

K2.5 Agent Swarm (Beta): Multi-agent parallel coordination

Beta Access: Agent Swarm with free credits for high-tier paid users

Platform Options

Kimi.com: Web-based interface with all four modes

Kimi App: Mobile/desktop application

API: platform.moonshot.ai for developer integration

Kimi Code: Terminal-based coding assistant

Open Source: Kimi Code released as open-source project

Configuration Details

Temperature: 1.0 (default)

Top-p: 0.95 (default)

Context Length: 256k tokens

Reproducibility: Official API recommended for benchmark recreation

Vendor Verification: Kimi Vendor Verifier (KVV) for third-party services

Part VII: The Road to AGI

Meaningful Step Forward

For Open-Source Community: Most powerful model demonstrating real-world capability

Under Real Constraints: Strong performance within practical limitations

Production Readiness: Suitable for actual knowledge work deployment

The Future Direction

Continued Advancement: Pushing further into agentic intelligence frontier

Boundary Redefinition: Challenging assumptions about AI capabilities in knowledge work

Research Focus: Expanding parallel coordination and visual reasoning

Open Ecosystem: Contributing to accessible AI advancement

Conclusion: Visual Agentic Intelligence Arrives

The Three Pillars

1. Coding with Vision: Native multimodal architecture enabling visual debugging and image-to-code workflows

2. Agent Swarm: Self-directed parallel coordination with up to 100 sub-agents and 1,500 tool calls

3. Office Productivity: Expert-level document/spreadsheet/PDF/slide generation through conversation

The Performance Story

59.3% improvement on AI Office Benchmark over K2 Thinking

24.3% improvement on General Agent Benchmark

4.5x speedup through agent swarm parallelization

State-of-the-art coding and vision capabilities

Fraction of cost compared to proprietary competitors

The Technical Innovation

15 trillion tokens of vision-text joint pretraining

PARL training with staged reward shaping

Critical-steps metric forcing genuine parallelism

No predefined workflows in agent orchestration

Autonomous visual debugging capability

The Accessibility

Open-source model pushing frontier forward

Multiple access points: Web, app, API, terminal

Four operational modes for different use cases

Beta features with free credits for experimentation

The Paradigm Shift

From sequential to parallel agent execution

From text-only to native multimodal reasoning

From hours to minutes for complex knowledge work

From predefined to self-directed workflow coordination

---

Get Started:

• Web: https://www.kimi.com
• API: https://platform.moonshot.ai
• Code: https://www.kimi.com/code
• Modes: Instant, Thinking, Agent, Agent Swarm (Beta)

Technical Report: Full details including prompts and methodology forthcoming

Vendor Verification: https://kimi.com/blog/kimi-vendor-verifier.html

WorldVQA Benchmark: https://github.com/MoonshotAI/WorldVQA

---

The Bottom Line: Kimi K2.5 represents the most powerful open-source model to date, achieving breakthrough performance through native multimodal architecture (15T vision-text tokens), self-directed agent swarm coordination (100 sub-agents, 1,500 tool calls, 4.5x speedup), state-of-the-art coding with vision (autonomous visual debugging), and expert-level office productivity (59.3% AI Office improvement, 24.3% General Agent improvement). The combination of visual agentic intelligence with PARL-trained parallel orchestration marks meaningful step toward AGI for open-source community, demonstrating strong capability on real-world tasks under real-world constraints at fraction of proprietary model costs. Access via Kimi.com, app, API, and open-source Kimi Code terminal tool across four modes (Instant/Thinking/Agent/Agent Swarm Beta). The future of agentic intelligence is parallel, visual, and open.

Try Agent Swarm Beta: Experience 100-agent coordination redefining knowledge work efficiency. 🦞✨