Kimi K2 for roleplay and creative writing: setup and an honest read.

Quick answer: three values, and kimi-k2.6 as the story id.

Any frontend with a custom OpenAI-compatible option takes the same three values: endpoint https://api.apisrouter.com/v1, your sk-... key, and the model id typed exactly. For roleplay and creative writing use kimi-k2.6, the general model; kimi-k2.7-code is a coding specialist and the wrong tool for prose even though it answers. The same connection serves DeepSeek, GLM, and Claude ids, so comparing K2's voice against the community defaults mid-story is a dropdown change, not a new account. Given how split the community verdict on K2 prose is, that comparison is not optional homework; it is the whole method.

Endpoint:  https://api.apisrouter.com/v1
API key:   sk-...            (from APIsRouter)
Model:     kimi-k2.6         (typed exactly; not kimi-k2.7-code, that one is for code)

What the evidence actually says about K2 prose.

The benchmark case is real. On EQ-Bench's creative-writing evaluation, the K2 line's instruct model scored high bands across the judged dimensions, adherence to instructions, nuanced characters, imagery, coherence, with elegant-prose and emotional-engagement marks just behind, and those public sample outputs are worth reading directly rather than taking anyone's summary. The community case is split, and honestly so. Plenty of roleplay users praise K2's distinctive, less sanitized voice and its willingness to keep a scene's energy. But named critics exist too: one widely shared 2026 write-up called the K2 Thinking generation a great agent and a mediocre writer, arguing that benchmark wins were being projected onto prose quality, and reporting coherence drop-off in demanding long-form work. Both things can be true: strong judged short-form output, and long-session drift that benchmarks do not measure. The practical read: K2 belongs on your shortlist, especially if the mainstream defaults feel same-y, and it needs to earn the seat in your own frontend against deepseek-v4-flash on price and claude-sonnet-4-6 on ceiling. Five long scenes on each answers more than any review, including this one.

SillyTavern setup, field by field.

SillyTavern appends /chat/completions itself, so the classic failure is pasting the full completions URL and getting a 404 from a doubled path; the field wants the base ending at /v1. An empty model dropdown after Connect means the /v1/models call failed: recheck the URL, the key, and that the source is Custom (OpenAI-compatible) rather than plain OpenAI. Use Chat Completion mode, not Text Completion, for any hosted endpoint. If chats fail after the model list works, run one curl against /v1/chat/completions with the same key; the error that comes back is the provider's real one rather than the frontend's generic message.

• Open API Connections (the plug icon).
• Set API to Chat Completion.
• Set Chat Completion Source to Custom (OpenAI-compatible).
• Custom Endpoint (Base URL): https://api.apisrouter.com/v1, stopping at /v1 with no trailing path.
• Custom API Key: your sk-... key.
• Click Connect; pick kimi-k2.6 from the dropdown that fills from /v1/models.

RisuAI setup.

In RisuAI, open Settings, go to the API tab, and pick the custom OpenAI-compatible provider rather than a named preset. Enter the key, the endpoint https://api.apisrouter.com/v1, and type kimi-k2.6 into the model field, enabling the custom-model option if the dropdown does not know the id; without that toggle the app quietly keeps the last built-in model selected, which is the most common silent misconfiguration. Labels vary slightly across RisuAI's web, desktop, and mobile builds, but the three values are the same everywhere. On the web build the browser itself makes the API call, so a config that looks right but produces silence is worth checking in the developer console for a CORS block; the desktop build sidesteps that class of problem. Save, open a character, send one short message. A reply proves endpoint, key, and model id in one step.

Settings that suit K2's voice, and the drift problem.

Community sampler lore for K2 is thinner than for the DeepSeek family, so start conservative and move one knob at a time. The bigger lever is structural: the criticism that sticks to this family is long-session coherence, and the mitigation is the same one that works everywhere, a stable card, a bounded working window, and summaries carrying older canon forward, so the model re-reads a clean compressed past instead of three hundred raw messages. K2.6 is also a reasoning-capable model, which shapes the feel: some replies pause before the first visible word while a thinking pass runs, and that pass bills as output tokens. If turns feel slow or bill heavier than the visible prose, cap response length first and judge again before blaming the model.

Context budgeting on a 256K window.

K2.6's 256K context, as served here, holds a detailed character card plus a very long visible history without forcing truncation. The constraint is financial rather than technical: frontends resend the whole allowed context every turn, so a story allowed to accumulate 100K tokens of history bills 100K input tokens per message. The economical shape is a deliberate working window, 32K carries a rich card and a long scene comfortably, with lorebook entries or summaries holding older canon. Spend the big window where it earns its cost: a finale that needs deep callbacks, or a reread, not every casual turn. This also directly addresses the drift criticism above, since a curated context is easier to stay coherent over than a raw one. At catalog rates the input side is what compounds; the pricing table below puts kimi-k2.6 next to the models you would compare it with, and the context-length guide linked at the bottom works the per-turn math across window sizes.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Content policy, stated plainly.

Kimi models operate under Moonshot's usage policies, and requests through APIsRouter carry the upstream model's content rules plus the terms of your frontend and any platform you publish on. Community discussion tends to describe the K2 family's fiction posture as comparatively flexible with a distinctive voice, and the accurate version of that observation is relative, a statement about where policy lines sit, not an absence of lines. Keep scenarios within each platform's rules and age settings, and check the policies-compared page linked below before building a long-running story on any single model's current posture, because provider policies are living documents.

FAQ