Kimi K2.5 Guide: Native Vision, Agent Swarm and Two Modes

Kimi K2.5 is the release where Moonshot's K2 line learned to see. Earlier K2 models were text-only; K2.5 was continually pretrained on images and text together, and it shipped with a headline feature that got more attention than the vision did: an agent swarm that can coordinate up to a hundred sub-agents.

The swarm is the flashier claim and the less practical one. The vision is the part most teams will actually use. This guide separates the two.

From text-only to native vision

K2.5 arrived in late January 2026; OpenRouter lists it from 27 January. Underneath it is the same mixture-of-experts shape as Kimi K2: 1T total parameters, 32B active, 384 experts with eight routed and one shared, and a 256K context window.

What is new is a MoonViT vision encoder of about 400M parameters and continual pretraining on roughly 15T mixed visual and text tokens, starting from Kimi-K2-Base. The model takes images as well as text, and Moonshot's announcement also shows video input for some workflows.

Joint pretraining is the detail that matters. A text model with a vision adapter attached afterwards tends to describe images well and reason about them poorly. Training on both from the base model up is the approach most likely to produce a model that can look at a screenshot and reason about what is wrong with it. Multimodal models explained covers how images become tokens in the first place.

Visual coding, read carefully

Moonshot leans hard on frontend work: turning a conversation into an interactive interface, generating code from images or video, reconstructing a website from a screen recording, and inspecting its own rendered output to iterate. Its model card reports 92.3 on OCRBench, 78.5 on MMMU-Pro and 86.6 on VideoMMMU.

Demos of generated web pages are the weakest kind of evidence, because they are chosen. The stronger evidence is the closed loop: render the page, screenshot it, hand the screenshot back, ask for the fix. That loop works with any vision-language model, and it is the right way to test this one. Best model for vision tasks covers what to put in that evaluation.

On text-only coding, Moonshot reports 76.8 on SWE-bench Verified and 50.8 on Terminal-Bench 2.0, the latter run in non-thinking mode. For comparison, K2 Thinking reported 71.3 on SWE-bench Verified, so K2.5 is an upgrade on code as well as gaining vision.

Agent swarm: what is real and what is a preview

The announcement describes an orchestrator that creates up to 100 sub-agents on the fly, coordinates up to 1,500 tool calls, and cuts execution time by up to 4.5 times compared with a single agent. Moonshot trained it with what it calls Parallel-Agent Reinforcement Learning, training the orchestrator while dispatching frozen sub-agents, to avoid the orchestrator falling back to running sub-agents one at a time.

The benchmark example is BrowseComp: 60.6 for a single agent, 74.9 with context management and 78.4 with the swarm. That is a real gain, and it comes on a task that splits naturally into independent searches.

Two caveats. First, Moonshot labelled the swarm a research preview and a beta in its own app, which is a different thing from a model capability you call through a plain completion endpoint. Second, parallelism only helps on work that actually parallelises. Most coding tasks have dependencies between steps, and a hundred sub-agents editing one repository is a coordination problem before it is a speed-up. Agent fan-out limits covers where width stops paying.

Instant or thinking: one model, two price profiles

K2.5 runs in a thinking mode or an instant mode, with thinking switched off by passing thinking set to disabled in the request body. Moonshot recommends temperature 1.0 for thinking and 0.6 for instant, with top-p 0.95 in both.

That switch is a cost control. On OpenRouter, K2.5 is listed at $0.50 input and $2.50 output per million tokens at the time of writing, with cached input at $0.15, and reasoning tokens are billed as output. Routing simple requests through instant mode and reserving thinking for hard ones can cut the output bill substantially without changing models. Model routing and fallbacks covers how to make that decision automatically.

The weights use the same native INT4 approach as K2 Thinking and ship under Moonshot's Modified MIT licence, which adds a requirement to display the Kimi K2.5 name above 100 million monthly active users or $20 million monthly revenue.

K2.5 or K2.6?

K2.6 followed in April 2026 with the same 1T/32B shape, the same 256K window and native vision, so the two are close relatives rather than different classes of model. The K2.6 guide covers its pricing and limits.

• Start with K2.6 for new work. It is the newer general model in the line, and the one Moonshot's current pricing page lists alongside K2.7 Code.
• Keep K2.5 if it already passes your evaluation and its listed price is lower for your traffic mix. Re-run the comparison rather than assuming the newer model wins on your tasks.
• Look at K2.7 Code if the job is purely programming. The K2.7 Code guide covers that specialist.

Whichever you test, include screenshots in the evaluation set. Vision is what K2.5 added to the line, and a text-only comparison will miss the reason you might pick it.