KwaiKAT KAT-Coder Models

KwaiKAT (Kwaipilot / Kuaishou StreamLake) KAT-Coder models are agentic coding models built for multi-step software work inside real repositories. They are called with the kwaikat/... canonical model id over /v1/chat/completions.

Model	Context	Input / Output / Cache Read ($/1M)	Note
kwaikat/kat-coder-pro-v2.5	256K	0.74 / 2.96 / 0.15	Flagship; repository-scale agentic work.
kwaikat/kat-coder-air-v2.5	256K	0.15 / 0.60 / 0.03	Cost-efficient; high-volume workflows.

Both models support tool/function calling, base64 image input, and prompt-cache reads.

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kwaikat/kat-coder-pro-v2.5",
    "messages": [
      { "role": "user", "content": "Run the tests in this repo and fix the failing one." }
    ],
    "max_tokens": 2048
  }'

Reasoning is always on

KAT-Coder models reason step by step on every request before answering; reasoning cannot be turned off and fields such as reasoning_effort are ignored. Reasoning tokens are returned inside completion_tokens and billed as output (not charged separately).

The practical consequence: if max_tokens is too low, the budget can be spent during reasoning and content comes back empty with finish_reason length. Keep max_tokens comfortable even for short answers (recommended: at least 1024). See Reasoning Effort for details.

Image input

The models accept images only as base64 data URLs (data:image/...;base64,...). Remote http(s) image URLs are not supported and the gateway rejects them with 400 unsupported_input.

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kwaikat/kat-coder-air-v2.5",
    "messages": [
      {
        "role": "user",
        "content": [
          { "type": "text", "text": "Summarize the error message in this screenshot." },
          { "type": "image_url", "image_url": { "url": "data:image/png;base64,<base64>" } }
        ]
      }
    ],
    "max_tokens": 1024
  }'

Tool calling

Standard OpenAI tools / tool_choice fields are supported. When the model calls a tool, finish_reason is tool_calls and content stays empty. See Tool Calling for the general flow.

Cache

When you resend the same prompt prefix, prompt-cache read kicks in; usage.prompt_tokens_details.cached_tokens is populated and those tokens are billed at the cache-read price. The cache warms up after a few requests. There is no cache-write fee.