Kimi Releases K2.5 Visual Agentic Model
In its February 2026 paper, the Kimi Team introduced Kimi K2.5, an open-source native multimodal agentic model, and released a post-trained checkpoint. Kimi's technical blog reports that the model can self-direct an Agent Swarm of up to 100 sub-agents and 1,500 parallel tool calls, with execution time reduced by up to 4.5x versus a single-agent setup.
Kimi has introduced Kimi K2.5, an open-source native multimodal agentic model, and released its post-trained checkpoint. The accompanying arXiv paper, revised on August 7, describes K2.5 as a model jointly optimized for text and vision through joint text-vision pretraining, zero-vision supervised fine-tuning, and joint text-vision reinforcement learning.
According to the Kimi technical blog, K2.5 continues pretraining from Kimi K2 on approximately 15 trillion mixed visual and text tokens. The company describes the release as supporting coding, visual reasoning, tool use, and an agent orchestration feature called Agent Swarm.
A multimodal MoE model
The public GitHub repository lists K2.5 as a mixture-of-experts model with 1 trillion total parameters and 32 billion activated parameters. Its published configuration includes 61 layers, 384 experts with eight selected per token, a 256K-token context length, Multi-head Latent Attention (MLA), and a 400M-parameter MoonViT vision encoder.
The paper frames its central technical objective as avoiding a trade-off between language and visual capability. It reports that K2.5 was trained using joint text-vision methods rather than treating vision as a separate downstream add-on. The authors report state-of-the-art results across coding, vision, reasoning, and agentic-task evaluations, though the abstract does not provide the underlying benchmark scores or evaluation protocols.
Kimi's blog highlights visual coding as a primary use case, including image- and video-to-code generation, visual debugging, interface generation, and visual data-processing workflows. These are company-reported demonstrations rather than independently reproduced results.
Agent Swarm orchestration
K2.5's most distinctive announced feature is Agent Swarm, which the paper describes as a self-directed parallel orchestration framework. According to Kimi, the system dynamically decomposes a complex request into heterogeneous subproblems, creates and coordinates agents without predefined workflows, and executes tasks concurrently.
The technical blog states that an Agent Swarm can use up to 100 sub-agents and make up to 1,500 tool calls for complex tasks. Both the blog and arXiv abstract report latency reductions of up to 4.5x relative to a single-agent baseline. Kimi.com offers Agent Swarm in beta, according to the blog, alongside Instant, Thinking, and Agent modes.
Parallel task decomposition can reduce wall-clock completion time where subtasks are genuinely independent, such as gathering sources, inspecting multiple files, or testing alternative code paths. In practice, practitioner evaluation will need to separate elapsed-time gains from total compute consumption: a system that launches many agents and tool calls can trade lower latency for higher inference, orchestration, and external-tool costs.
What practitioners can verify
The GitHub repository makes K2.5's architectural configuration available, while the arXiv paper supplies the claimed training and evaluation framing. The model's public checkpoint creates an opportunity for teams to test several operational questions that benchmark summaries alone do not settle:
- •Whether visual inputs improve coding or debugging accuracy on an organization's own UI, video, and documentation artifacts.
- •How reliably the swarm decomposes tasks, handles dependencies, and recovers when individual sub-agents or tools fail.
- •The cost, rate-limit, observability, and security implications of workflows that can generate large numbers of concurrent tool calls.
- •Whether long-context and multimodal workloads fit available GPU memory, serving infrastructure, and latency targets.
The release is notable because it combines a large open model, native vision-language training, and a public implementation of parallel agent orchestration. Claims of leading benchmark performance and 4.5x latency reductions remain claims from the Kimi paper and technical blog; independent testing will be necessary to establish reproducibility across tool environments and real production workloads.
Key Points
- 1Kimi released an open multimodal checkpoint, giving researchers a 1T-parameter MoE model with 32B activated parameters to evaluate.
- 2Kimi reports Agent Swarm can orchestrate 100 sub-agents and 1,500 tool calls, shifting evaluation beyond single-agent benchmark accuracy.
- 3Comparable parallel-agent systems can reduce elapsed time, but practitioners should measure total compute, tool costs, failure recovery, and observability.
Scoring Rationale
K2.5 is a substantial open model release combining native multimodality, a 1T-parameter MoE architecture, and a public agent-orchestration approach. Its checkpoint and reported swarm capability are highly relevant to teams evaluating open agentic systems, although major performance and latency claims still require independent reproduction.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
