BeamWeaver Moonshot/Kimi

BeamWeaver includes an OpenAI-compatible Moonshot provider under BeamWeaver.Moonshot for Kimi chat models.

Implemented

BeamWeaver.Moonshot.ChatModel implements Moonshot Chat Completions through BeamWeaver.Core.ChatModel.
Model initialization uses explicit Moonshot identifiers: BeamWeaver.Models.init_chat_model("moonshot:kimi-k3").
Bare kimi-* and kimi:* identifiers are rejected so provider routing stays explicit.
Defaults load from config :beam_weaver, :moonshot; runtime config reads MOONSHOT_API_KEY and optional MOONSHOT_BASE_URL or MOONSHOT_API_URL.
Chat requests support text, image, and video input; data: media URLs and Moonshot ms://... media references are accepted.
K3 always reasons and uses the top-level reasoning_effort: "max" option; it rejects the K2.x thinking object. K2.7 Code models require thinking to stay enabled; K2.6 and K2.5 accept thinking: %{type: "enabled" | "disabled"}. reasoning_content is preserved in message content blocks and normalized response metadata.
Standard function tools use the OpenAI-compatible tools shape. K3 accepts tool_choice: "required" and dynamically loaded tools through contentless system messages created by BeamWeaver.Moonshot.Tools.dynamic_message/1. Kimi $web_search is available through BeamWeaver.Moonshot.Tools.web_search/1 on K2.6/K2.5 and requires thinking: %{type: "disabled"}.
Structured output supports json_object and json_schema response formats.
Partial mode is emitted from assistant message metadata partial: true; JSON object mode with partial mode fails before transport.
Streaming supports text deltas, reasoning deltas, tool-call chunks, usage nested under either the response or final choice, reconstructed final assistant messages, and typed stream events.
Token counting uses /v1/tokenizers/estimate-token-count.
Checked-in profiles include kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed, kimi-k2.6, and kimi-k2.5; discontinued Kimi slugs fail before transport with K3 replacement metadata.

Usage

model =
  BeamWeaver.Moonshot.chat_model(
    model: "kimi-k3",
    reasoning_effort: "max"
  )

BeamWeaver.Core.ChatModel.invoke(model, [
  BeamWeaver.Core.Message.user("Summarize the trace.")
])

Use model-string routing when constructing through the generic initializer:

{:ok, model} = BeamWeaver.Models.init_chat_model("moonshot:kimi-k3")

Require a tool call on K3:

BeamWeaver.Core.ChatModel.invoke(model, messages,
  tool_choice: :required,
  tools: [weather_tool]
)

Dynamically load a tool at a specific point in K3 conversation history:

messages = [
  BeamWeaver.Core.Message.user("Calculate 23 * 47"),
  BeamWeaver.Moonshot.Tools.dynamic_message([calculator_tool])
]

BeamWeaver.Core.ChatModel.invoke(model, messages)

Use Kimi web search with a K2.6/K2.5 model and thinking disabled:

{:ok, model} = BeamWeaver.Models.init_chat_model("moonshot:kimi-k2.6")

BeamWeaver.Core.ChatModel.invoke(model, messages,
  thinking: %{type: "disabled"},
  tools: [BeamWeaver.Moonshot.Tools.web_search()]
)

Current Model Policy

kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed, kimi-k2.6, and kimi-k2.5 are supported Moonshot profiles. K3 has a 1,048,576-token context window, a 131,072-token default completion limit, and a 1,048,576-token maximum completion limit. It always reasons, currently accepts only reasoning_effort: "max", supports tool_choice values auto, none, and required, and is the only Kimi model that accepts dynamic tool messages. K2.7 Code models are thinking-only and accept only automatic tool choice ("auto" or "none") while thinking is enabled. K2.6 and K2.5 support both thinking and non-thinking modes; fixed sampling values are validated before transport.

K3 accepts text plus base64 or ms:// image/video input, streaming, parallel function calls, JSON object mode, strict JSON Schema output, Partial Mode, automatic prefix caching, prompt cache keys, safety identifiers, and token estimation. Sampling is fixed at temperature=1.0, top_p=0.95, n=1, and zero presence/frequency penalties. Complete assistant messages, including reasoning_content and tool_calls, must be replayed unchanged in later turns.

Older Kimi slugs such as kimi-latest, kimi-thinking-preview, kimi-k2-0905-preview, kimi-k2-0711-preview, kimi-k2-turbo-preview, kimi-k2-thinking, and kimi-k2-thinking-turbo are rejected with :deprecated_model and moonshot:kimi-k3 as the replacement. K2.5 is unavailable to newly registered Kimi users and is scheduled for full sunset on August 31, 2026.

Moonshot Files, Batch, Balance, and Formula/Fiber tool APIs are not exposed in BeamWeaver yet. Chat can still reference provider-hosted media with documented ms://... image or video URLs.

Kimi currently marks its web-search integration as being updated and advises against near-term production use. BeamWeaver therefore keeps the legacy $web_search helper limited to models where thinking can be disabled and does not advertise it as a K3 capability.