Documentation

WebikAI Inference is a provider-only API serving one open-weight model, kwaipilot/kat-coder-v2.5-dev, behind the standard OpenAI Chat Completions surface at https://api.webik.ai/v1. Authenticate with a bearer key (wk_live_...). Every response names the immutable release that served it.

Quickstart

Stock OpenAI SDKs work unmodified — set base_url, nothing else. There is no WebikAI SDK to install and no custom client to maintain.

curl https://api.webik.ai/v1/chat/completions \
  -H "Authorization: Bearer wk_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kwaipilot/kat-coder-v2.5-dev",
    "messages": [{"role": "user", "content": "Write a binary search in Rust"}]
  }'

from openai import OpenAI

client = OpenAI(base_url="https://api.webik.ai/v1", api_key="wk_live_...")

completion = client.chat.completions.create(
    model="kwaipilot/kat-coder-v2.5-dev",
    messages=[{"role": "user", "content": "Write a binary search in Rust"}],
)
print(completion.choices[0].message.content)
print(completion.system_fingerprint)  # wbk_... — the exact serving release

import OpenAI from 'openai';

const client = new OpenAI({
  baseURL: 'https://api.webik.ai/v1',
  apiKey: 'wk_live_...'
});

const completion = await client.chat.completions.create({
  model: 'kwaipilot/kat-coder-v2.5-dev',
  messages: [{ role: 'user', content: 'Write a binary search in Rust' }]
});
console.log(completion.choices[0].message.content);

Models and releases

Three layers of identity, from stable to exact:

• Canonical ID — kwaipilot/kat-coder-v2.5-dev, the name you request. Points at the current stable release.
• Release ID — an immutable, dated build, e.g. webik/kat-coder-v2.5-dev-2026-08-w4a16-r1 (an example of the form). Fixes weights, quantization, engine, chat template, tool parser, and context configuration.
• Fingerprint — system_fingerprint (wbk_ + 12 hex) on every response, unique per release; the X-Webik-Model-Release header carries the full release ID alongside it.

GET /v1/models lists the canonical ID; GET /v1/model-releases lists releases with their status, dates, and any deprecation schedule:

curl https://api.webik.ai/v1/models -H "Authorization: Bearer wk_live_..."
curl https://api.webik.ai/v1/model-releases -H "Authorization: Bearer wk_live_..."

Pin a release either way — the response echoes exactly the model string you sent:

// Pin by sending the release ID as the model:
{ "model": "webik/kat-coder-v2.5-dev-2026-08-w4a16-r1", "messages": [ ... ] }

// Or keep the canonical ID and pin via the webik.release extension:
{
  "model": "kwaipilot/kat-coder-v2.5-dev",
  "webik": { "release": "webik/kat-coder-v2.5-dev-2026-08-w4a16-r1" },
  "messages": [ ... ]
}

Full release semantics and the deprecation policy (90-day overlap, 30-day notice, Deprecation/Sunset headers) live on the model page.

Streaming

Set stream: true for server-sent events. Ask for stream_options.include_usage and the final chunk carries the same engine-reported usage billing settles from — your numbers and our numbers are the same numbers.

stream = client.chat.completions.create(
    model="kwaipilot/kat-coder-v2.5-dev",
    messages=[{"role": "user", "content": "..."}],
    stream=True,
    stream_options={"include_usage": True},
)
for chunk in stream:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")
    if chunk.usage:  # final chunk — what billing settles from
        print(chunk.usage)

Tool calling

The standard tools array and tool_choice parameter work as documented by OpenAI. KAT-Coder is trained for agentic tool use; parallel tool calls are supported, and tool-call parsing is part of the release contract — a parser change mints a new release, so an agent that works today cannot silently break tomorrow.

{
  "model": "kwaipilot/kat-coder-v2.5-dev",
  "messages": [{"role": "user", "content": "Run the failing test, then fix it"}],
  "tools": [{
    "type": "function",
    "function": {
      "name": "run_tests",
      "description": "Run the project test suite",
      "parameters": {
        "type": "object",
        "properties": {"filter": {"type": "string"}}
      }
    }
  }]
}

Structured output

response_format supports json_object (any valid JSON) and json_schema (constrained decoding against your schema):

{
  "model": "kwaipilot/kat-coder-v2.5-dev",
  "messages": [{"role": "user", "content": "Extract the dependencies from this Cargo.toml"}],
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "dependencies",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "dependencies": {"type": "array", "items": {"type": "string"}}
        },
        "required": ["dependencies"]
      }
    }
  }
}

Reasoning effort

reasoning_effort accepts none, medium, or high and controls how much the model deliberates before answering. Both forms below are Webik extensions resolved at the edge and stripped before the request reaches the engine — the engine never sees a foreign field.

// OpenAI-style top-level parameter:
{ "model": "kwaipilot/kat-coder-v2.5-dev", "reasoning_effort": "high", "messages": [ ... ] }

// Equivalent nested form for clients that reject unknown top-level fields:
{ "model": "kwaipilot/kat-coder-v2.5-dev", "webik": { "mode": "high" }, "messages": [ ... ] }

Errors

Errors use the OpenAI envelope with a stable machine code in error.code. Capacity errors (429) carry retry_after_seconds in the body and never debit the wallet — a failed request costs nothing. Match on error.code, not the HTTP status: capacity refusals share 429 with rate_limit_exceeded.

Billing

Billing is a prepaid wallet ($20 minimum top-up via Stripe Checkout). Before dispatch, a conservative reservation is held against your balance; after the response, the request settles to engine-reported usage at $1.00 per 1M input and $4.00 per 1M output tokens (introductory), and the unused remainder is released. A small dynamic minimum request fee covers cold starts. The price version in force is recorded on every usage row, and your history is exportable as CSV.

Data handling

No prompt or response body storage by default. Request logs are metadata only: timestamps, token counts, release ID, price version, status, and request ID. Bodies pass through to the inference engine and are not persisted by WebikAI. Details in the privacy policy.

Use it from your tools

Anything that speaks the OpenAI API can point at https://api.webik.ai/v1. Setup guides for the coding agents the beta targets:

• Claude Code
• Cursor
• Cline
• Aider
• OpenCode