LLMTR Gemma 4 (Turkey)

Gemma 4 is a first-party model we host on our own hardware in Turkey. It is called through /v1/chat/completions with the canonical model id llmtr/gemma-4. Requests are never forwarded to a third-party provider, which makes it a fit for internal assistants, document analysis and tool-using agent flows with data-residency requirements.

Property	Value
Model id	`llmtr/gemma-4`
Underlying model	Gemma 4 26B-A4B-IT
Context window	131,072 tokens (128K)
Price (input / output / cache read)	$2.00 / $5.00 / $0.50 per 1M tokens
Image input	Yes (base64 data URLs only)
Tool calling	Yes
Prompt cache	Yes
Reasoning	Yes, optional — off by default
Audio / video input	No
Image generation	No

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llmtr/gemma-4",
    "messages": [
      { "role": "user", "content": "Summarise this contract clause in plain English." }
    ],
    "max_tokens": 1024
  }'

Reasoning is optional

By default the model answers without thinking, so no reasoning tokens are spent. For tasks that need step-by-step reasoning you can turn it on in either of two ways:

{ "model": "llmtr/gemma-4:think", "messages": [{ "role": "user", "content": "Solve this logic puzzle." }] }
{ "model": "llmtr/gemma-4", "reasoning": true, "messages": [{ "role": "user", "content": "Solve this logic puzzle." }] }

Use the :fast selector or reasoning: false to turn it off. When reasoning is on, the trace is returned in message.reasoning_content, is spent from your max_tokens budget and is billed as output. With a very small max_tokens the budget can be consumed by reasoning alone and content may come back empty.

Image input

The model reads images, but accepts them only as base64 data URLs. Sending a remote https:// address returns a 400 unsupported_input.

{
  "model": "llmtr/gemma-4",
  "messages": [
    {
      "role": "user",
      "content": [
        { "type": "text", "text": "What is in this image?" },
        {
          "type": "image_url",
          "image_url": { "url": "data:image/png;base64,iVBORw0KGgo..." }
        }
      ]
    }
  ],
  "max_tokens": 1024
}

Image tokens are billed as ordinary input tokens; there is no separate image rate.

Tool calling

The model supports the standard OpenAI tools schema and returns real tool_calls:

{
  "choices": [
    {
      "message": {
        "role": "assistant",
        "content": "",
        "tool_calls": [
          {
            "id": "call_1",
            "type": "function",
            "function": { "name": "get_weather", "arguments": "{\"city\":\"Istanbul\"}" }
          }
        ]
      },
      "finish_reason": "tool_calls"
    }
  ]
}

Prompt cache

Repeated prompt prefixes are served from cache. Cached tokens are reported in prompt_tokens_details.cached_tokens and billed at the cache read rate instead of input:

"usage": {
  "prompt_tokens": 1603,
  "completion_tokens": 2,
  "prompt_tokens_details": { "cached_tokens": 1579 }
}

Here 1,579 tokens are charged at $0.50/1M and the remaining 24 at $2.00/1M.

Speed and long prompts

Generation runs at roughly 30-45 tokens per second, and the 128K context window is usable in practice for long documents. On the client side:

For long outputs use stream: true and keep your client read timeout at 120 seconds or more. The 60-second default in most HTTP clients is not enough for long answers.
Resend the same prefix (system instructions, document, conversation history): the prefix is cached, responses get noticeably faster, and cached tokens are billed at the cheaper rate.
Response times can grow during busy periods, so do not keep read timeouts tight.

See Timeouts and slow models for details.