llm-api-adapter-kimi

An optional Kimi / Moonshot organization package for LLM API Adapter, a Python SDK with one shared interface for calling LLM APIs.

Calls Kimi / Moonshot through POST /v1/chat/completions without a provider SDK.

Installation

Install through the Core package extra:

pip install "llm-api-adapter[kimi]"

Direct installation is also supported when Core is managed separately:

pip install llm-api-adapter-kimi

Async requests need HTTPX:

pip install "llm-api-adapter[kimi,async]"

Synchronous requests remains the default. To opt into HTTPX for sync chat() and stream_chat(), install "llm-api-adapter[kimi,httpx]" and pass transport="httpx".

Quick start

import os

from llm_api_adapter.models.messages.chat_message import UserMessage
from llm_api_adapter.universal_adapter import UniversalLLMAPIAdapter

adapter = UniversalLLMAPIAdapter(
    organization="kimi",
    model="kimi-k3",
    api_key=os.environ["KIMI_API_KEY"],
)

response = adapter.chat(
    messages=[UserMessage("Explain retrieval-augmented generation.")],
    max_tokens=128,
)
print(response.content)

Supported models

• kimi-k3
• kimi-k3.1-pro
• kimi-k3.1-pro-highspeed
• kimi-k2.6

Capabilities

reasoning_level is resolved automatically from registry metadata. K3 and K3.1 Pro cannot disable reasoning, so reasoning_level="none" uses the registered minimum and warns. K2.6 maps "none" to disabled thinking and every other valid level to enabled thinking. When omitted, no thinking control is sent and Kimi's native default is preserved. Reasoning is never mixed into visible text; use capture_reasoning=True for opt-in observability.

All models reject named tool choice before transport. kimi-k3 and the K3.1 Pro models accept tool_choice="any"; kimi-k2.6 rejects it and accepts only automatic or disabled tool choice. Exact supported values, request rules, and exceptions are recorded in the Kimi registry.

History and files

Kimi Chat Completions is stateless: previous_response is accepted for the shared API but is not serialized. Send the complete messages history on each turn, including assistant tool calls and ToolMessage results.

Image bytes are encoded as data URIs. Public image URLs and every DocumentPart are rejected before transport. Kimi's Files API provides extracted text rather than a Chat Completions attachment, so the adapter does not upload, retain, download, extract, or delete caller files.

Kimi maps authentication/authorization (401/403), rate-limit (429), timeout (408/504), documented token/quota, and server failures to the matching public LLMAPI*Error; other client or SSE failures become LLMAPIClientError.

Automatic cache usage and pricing

usage.prompt_tokens_details.cached_tokens becomes Usage.cached_tokens; the legacy usage.cached_tokens field is also accepted. K3 and K3.1 Pro additionally report prompt_tokens_details.cache_write_tokens as Usage.cache_write_tokens for their automatic cache-write component. Both are disjoint subsets of total Usage.input_tokens. K2.6 has registered automatic cache-read pricing only.

Input cost uses ordinary, read, and applicable write rates from the USD registry tier without double counting. If a required read or write count is omitted, cost_input and cost_total stay None; known output cost can still be exposed. An explicitly reported zero stays zero. The adapter does not enable opt-in cache modes or account for selectable TTLs or storage charges. See the shared usage and pricing guide.

See the main llm-api-adapter README for the shared API contract, error mapping, and E2E/release documentation.