Kimi K2.5

Seeing changes what an agent can fix

K2.5 extends K2 with a vision encoder, a projection that maps the encoder’s output into the language model’s representation, and training on images and text together. The report’s encoder, MoonViT-3D, represents images and video in the same space. For video, after the encoder has processed the frames, it averages the corresponding patch representations across groups of up to four consecutive frames before passing them to the language model. That cuts the number of video tokens by up to four times, at the cost of some detail about motion. The language model itself is still the sparse K2 design. [1]

Imagine fixing a web page from a screenshot. The source code may say that two labels are centered, but the rendered page shows them overlapping. A model that sees the rendering can propose a different fix from one that reads only the source. Seeing helps only if the model spots the actual problem instead of offering generic design advice.

The report uses the phrase “zero-vision SFT” in a narrow sense. After pretraining on images and text, a supervised fine-tuning stage that uses only text is enough to switch on the model’s tool-use behavior. It does not mean the model learned to see without images. Reinforcement learning on visual tasks, rewarded by whether the answer is correct, comes afterward. [1]

Independent work can overlap

Suppose four independent lookups take 3, 2, 4 and 1 time units. One worker needs 10 units to do them all. With more workers the tasks can overlap, but the job is done only when the busiest worker finishes. The figure’s scheduler gives each new task to whichever worker frees up first; with two workers, the job takes 6 units. Change the number of workers to see the schedule.

This calculation says nothing about how capable the model is. It assumes the durations are known, the tasks are independent, nothing fails and coordination is free. A real agent first has to find a useful way to split the work. If task B needs task A’s result, giving them to different agents does not make them independent.

What sets the finishing time is the critical path: the longest chain of steps in which each depends on the one before. Adding agents helps only if it shortens that chain by more than the coordination it adds.

Rewarding useful delegation

K2.5’s Parallel-Agent Reinforcement Learning trains the orchestrator while the helper agents stay fixed. Early in training, extra rewards encourage the orchestrator to delegate and reward helpers for finishing their subtasks. These extra rewards are gradually reduced to zero, so that in the end only success on the whole task counts. [2]

To measure speed, the authors count critical steps. In each stage, they add the orchestrator’s own steps to the steps of the longest-running helper, then sum over the stages. If the orchestrator takes one step and three helpers then take 3, 2 and 4 steps in parallel, the stage costs 1 + 4 = 5 critical steps, although 10 steps were taken in total. These count model and tool steps, not seconds. [2]

The reward design guards against two bad strategies. An orchestrator might do everything itself because it never tries delegating. Or it might spawn many helpers simply because spawning is rewarded. A good reward separates real progress from visible activity.

In the lookup example, asking three helpers to fetch the same document does not produce three independent facts. Asking one to fetch a number, another to check its definition and a third to confirm its date range can help, but only if the orchestrator then reconciles what they found. Running in parallel makes disagreements appear at once; resolving them is still work.

The released model versus the full system

Moonshot’s launch describes Agent Swarm, the multi-agent mode, as a research preview in its own product. The open model card provides the K2.5 weights and instructions for running them. [2] [3] Downloading the weights does not give you the orchestrator’s training environment, the tools or the hosted swarm service.

When reading a reported agent score, check the tools, the thinking mode, how context was handled and how many attempts were allowed. The release’s evaluation notes vary these from task to task; some coding results use the non-thinking mode. One model can therefore appear inside several quite different systems. [2]

Both ideas on this page concern evidence. Vision changes what an agent can observe. Parallel helpers change how quickly it can gather independent observations. Neither removes the need to check that the final answer actually follows from what was observed.

Simplifications and limits

• Agent Swarm results come from Moonshot’s hosted system; downloading the weights alone does not reproduce them.
• The scheduler adds up made-up task durations; it does not call any model or simulate reasoning.

Follow the idea

Kimi K2 ThinkingQwen3.5Kimi K3

Sources and further reading

Primary papers, model cards, and release posts. Facts describe the versions named here; the interactive calculations are teaching models. Research cutoff: October 3, 2026.

1. Kimi K2.5 technical reportSections 2–4; zero-vision SFT, visual RL, MoonViT-3D and PARL.
2. Kimi K2.5 official technical blogPARL rewards and critical steps; Agent Swarm product scope; evaluation footnotes.
3. Kimi-K2.5 model cardOpen checkpoint, architecture summary, inference and license.