MarsCompute

MarsCompute serves open models through an OpenAI-compatible API. The LLMTR catalog has two MarsCompute models; both are called with the marscompute/ prefix on the /v1/chat/completions endpoint. Both read and write text only, and thinking is switched off.

Model	Context	Input / Output ($/1M)
marscompute/mimo-v2.6-flash-marscompute	1,000,000	0.15 / 0.30
marscompute/qwen-sea-lion-v4.5-27b	8,192	0.60 / 3.60

There is no cache discount: all input is billed at the input price in the table.

Chat

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "marscompute/qwen-sea-lion-v4.5-27b",
    "messages": [
      { "role": "system", "content": "Answer briefly." },
      { "role": "user", "content": "Apa ibu kota Indonesia?" }
    ],
    "max_tokens": 512
  }'

Both models support streaming. Without max_tokens a reply is at most 4,096 tokens; send max_tokens for longer replies.

These rows have no tool calling and no JSON mode through response_format. A request that sends tools or a JSON response_format is refused with 400 invalid_request before it is sent; describe the output format in the prompt instead. A request with an image is refused too.

MiMo V2.6 Flash

marscompute/mimo-v2.6-flash-marscompute is Xiaomi's MiMo V2.6 Flash. It offers a context of about one million tokens and answers directly, without thinking.

This row does not support stop: a request that sends stop is refused with 400 invalid_request. Cut the reply at your own delimiter, or limit its length with max_tokens.

Qwen-SEA-LION-v4.5-27B-IT

marscompute/qwen-sea-lion-v4.5-27b is AI Singapore's instruction model for Southeast Asian languages, trained on top of Qwen3.6-27B. Alongside English it supports Burmese, Indonesian, Filipino, Malay, Tamil, Thai and Vietnamese. Turkish is not among its supported languages.

Prompt and reply together can be at most 8,192 tokens. For long prompts, set max_tokens so that it and the prompt stay within 8,192. The prompt text also cannot exceed about 32,000 characters; a longer prompt is refused with 400 invalid_request and is not billed.