Reading the following paper:

Kimi K2.5 is a multimodal agentic model designed to bridge the gap between vision-language models (VLMs) and autonomous agents. Built upon the Kimi K2 Mixture-of-Experts (MoE) architecture, K2.5 introduces two primary technical innovations: native multimodal joint optimization (enhancing both text and vision through co-training) and Agent Swarm, a framework for parallel agent orchestration. The model demonstrates state-of-the-art performance across reasoning, coding, and video understanding benchmarks, with the Agent Swarm framework reducing inference latency by up to 4.5x compared to sequential baselines.

1. Model Architecture & Infrastructure

  • Backbone: The model is based on Kimi K2, a 1.04 trillion parameter MoE (32B activated parameters per token) trained on 15 trillion tokens.
  • Vision Encoder (MoonViT-3D):
    • Utilizes the NaViT packing strategy, allowing variable-resolution image inputs without complex splitting.
    • Implements a lightweight 3D ViT compression mechanism where consecutive frames are grouped in fours and temporally averaged at the patch level. This provides 4x temporal compression, enabling the processing of videos up to 2,000+ frames within the context window.
  • Decoupled Encoder Process (DEP): To address load imbalances caused by variable multimodal inputs (e.g., varying image counts), the training infrastructure decouples the vision encoder from the main backbone. This allows the system to maintain 90% of the training efficiency of pure text training.

2. Key Technical Innovations in Training

A. Native Multimodal Pre-Training (Counter-Intuitive Finding)

Contrary to the prevailing “late fusion” consensus (where vision tokens are added at high ratios late in training), K2.5 employs early fusion with lower vision ratios. Ablation studies revealed that introducing vision tokens early at a moderate ratio (e.g., 10-20%) yields better multimodal and textual performance under a fixed token budget than aggressive late-stage injection, avoiding the “representation shock” that disrupts linguistic capabilities.

B. Zero-Vision SFT

During supervised fine-tuning (SFT), the team discovered that text-only SFT effectively activates visual tool-use capabilities.

  • Method: Image manipulations are proxied through programmatic operations (e.g., IPython) in text data.
  • Insight: Adding human-designed visual trajectories during SFT actually hurt generalization. Text-only SFT acts as a “cold start” mechanism, likely because the joint pre-training has already established strong alignment. This allows the model to perform pixel-level operations (like counting or binarization) without explicit vision-SFT data.

C. Joint Multimodal Reinforcement Learning (RL)

The model utilizes a Unified Agentic RL environment with Generative Reward Models (GRMs) that evaluate trajectories based on values like helpfulness and strict instruction following, rather than just binary outcomes.

  • Cross-Modal Transfer: A critical finding is that visual RL enhances textual performance. After visual RL, benchmarks like MMLU-Pro and GPQA-Diamond saw improvements (+1.7% and +2.1% respectively), suggesting that visual grounding reduces uncertainty in complex reasoning tasks.
  • Toggle Mechanism: To prevent “length-overfitting” (where models fail to generalize to higher compute scales), the training alternates between budget-constrained optimization and standard inference-time scaling phases. This reduces output tokens by 25-30% without performance degradation.

3. Agent Swarm: Parallel Orchestration Framework

To solve the linear latency scaling of sequential agents, K2.5 introduces Agent Swarm, a Parallel-Agent Reinforcement Learning (PARL) framework.

  • Architecture: A trainable Orchestrator dynamically instantiates and schedules frozen sub-agents.
    • Decoupling: Sub-agents are frozen during training to solve credit assignment ambiguity and training instability.
    • Context Management: Sub-agents maintain local contexts, returning only relevant results to the orchestrator. This acts as a “proactive context management” system, superior to reactive truncation strategies like “Discard-all”.
  • Reward Formulation: The PARL reward includes specific terms to guide behavior:
    • $r_{parallel}$: Incentivizes concurrent scheduling to prevent “serial collapse” (defaulting to single-agent mode).
    • $r_{finish}$: Prevents “spurious parallelism” (creating useless sub-agents to hack the parallelism reward).
  • Critical Steps Metric: Optimization focuses on minimizing Critical Steps—defined by the longest path in the execution graph (time taken by the main agent + the longest-running parallel sub-agent)—rather than total steps.

4. Evaluation Highlights

  • Reasoning: K2.5 achieves 96.1% on AIME 2025 and 87.1% on MMLU-Pro, performing competitively with GPT-5.2 and Claude Opus 4.5.
  • Coding: Scores 76.8% on SWE-Bench Verified and 85.0% on LiveCodeBench, demonstrating robustness in real-world software engineering.
  • Video Understanding: Establishes new SOTA results on VideoMMMU (86.6%) and LongVideoBench (79.8%), driven by the MoonViT-3D architecture.
  • Agentic Search: The Agent Swarm framework shines here, achieving 78.4% on BrowseComp (vs. 60.6% for single-agent K2.5) and reducing execution time by 4.5x on WideSearch tasks.

Takeaways

The most significant takeaway from Kimi K2.5 is the successful inversion of standard VLM training wisdom. By validating that early, low-ratio vision fusion outperforms late fusion, and that text-only SFT can bootstrap visual agents better than visual SFT, strong cross-modal alignment is foundational rather than additive. Furthermore, the Agent Swarm represents a practical leap in deploying agents for heavy workloads (e.g., 40GB video analysis), moving from “chain-of-thought” to “graph-of-agents” where latency is managed via learned parallelism rather than heuristics.