PublicAI Apertus

Apertus models are called with the publicai/... canonical model id. Three models from the Apertus v1.5 family are in the public catalog:

Model	Context	Image input	Tool calling	Reasoning	Input / Output (1M tokens)
publicai/apertus-v1.5-8b	262,144	Yes	Yes	No	$0.10 / $0.20
publicai/apertus-v1.5-8b-thinking	262,144	Yes	No	On by default, can be turned off	$0.10 / $0.20
publicai/apertus-v1.5-70b	262,144	Yes	Yes	No	$0.82 / $2.92

All three take text and image input, produce text output, support JSON mode, and offer a 262,144-token context window.

The previous-generation publicai/apertus-8b-instruct and publicai/apertus-70b-instruct models keep working. Their context window is 65,536 tokens and they do not accept image input; use the v1.5 versions for new integrations.

Chat Completions

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "publicai/apertus-v1.5-70b",
    "messages": [
      { "role": "user", "content": "What are the official languages of Switzerland?" }
    ],
    "max_tokens": 600
  }'

Turning reasoning on and off

Reasoning is only available on publicai/apertus-v1.5-8b-thinking, where it is on by default. The reasoning trace is not mixed into the answer; it comes back separately in message.reasoning_content.

To turn it off, append :fast to the model id:

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "publicai/apertus-v1.5-8b-thinking:fast",
    "messages": [
      { "role": "user", "content": "What is 2 + 2?" }
    ],
    "max_tokens": 200
  }'

The reasoning body field achieves the same result:

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "publicai/apertus-v1.5-8b-thinking",
    "messages": [
      { "role": "user", "content": "What is 2 + 2?" }
    ],
    "reasoning": false,
    "max_tokens": 200
  }'

The :think suffix keeps reasoning on and matches the default behaviour. With reasoning on, output token consumption rises noticeably, and reasoning tokens are billed as output. For short, single-step requests :fast lowers both latency and cost.

Turning reasoning off applies only to the -thinking model. The other two models do not support reasoning; do not send a reasoning suffix to them.

Tool calling

publicai/apertus-v1.5-8b and publicai/apertus-v1.5-70b support the OpenAI-compatible tools schema:

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "publicai/apertus-v1.5-8b",
    "messages": [
      { "role": "user", "content": "What is the weather in Ankara?" }
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_weather",
          "description": "Get the current weather for a city",
          "parameters": {
            "type": "object",
            "properties": { "city": { "type": "string" } },
            "required": ["city"]
          }
        }
      }
    ]
  }'

publicai/apertus-v1.5-8b-thinking does not support tool calling. Sending tools or tool_choice to that model is rejected with 400 unsupported_input. Use publicai/apertus-v1.5-8b for tool-calling flows.

See Tool calling for the general tool-calling flow.

Image input

All three v1.5 models read images. Images can only be supplied as base64 data URLs:

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "publicai/apertus-v1.5-70b",
    "messages": [
      {
        "role": "user",
        "content": [
          { "type": "text", "text": "What is in this image?" },
          {
            "type": "image_url",
            "image_url": { "url": "data:image/png;base64,BASE64_CONTENT" }
          }
        ]
      }
    ],
    "max_tokens": 400
  }'

Remote https:// image URLs are not supported and are rejected with 400 unsupported_input. Download the image and send it base64-encoded. Image tokens are reported inside usage.prompt_tokens and billed at the normal input token price; there is no separate image charge.

See Images for the general image flow.

JSON mode

All three models produce JSON output with response_format:

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "publicai/apertus-v1.5-8b",
    "messages": [
      { "role": "user", "content": "Return JSON with the city name and country for Ankara." }
    ],
    "response_format": { "type": "json_object" },
    "max_tokens": 200
  }'

Prompt cache

Apertus models have no prompt cache. Responses do not include a usage.prompt_tokens_details field and all input tokens are billed at the normal price.

Data policy

Requests are processed on PublicAI / Swiss AI infrastructure. These are not Turkey-hosted LLMTR models. Take that into account for production flows that carry sensitive data.