Gemma 4 26B A4B

Gemma 4 26B A4B is Google's open-weight Mixture-of-Experts model from the Gemma 4 family: a sparse architecture that activates only 3.8 billion of its 25.2 billion parameters per token, giving it fast inference alongside much larger model capacity. It is called through /v1/chat/completions with the canonical model id gemma/gemma-4-26b-a4b-it.

Model	Context	Input / Output ($/1M)
gemma/gemma-4-26b-a4b-it	131,072	0.042 / 0.22

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma/gemma-4-26b-a4b-it",
    "messages": [
      { "role": "user", "content": "Write a short joke about saving RAM." }
    ],
    "max_tokens": 512
  }'

Reasoning is off by default

The model can reason step by step before answering, but this is off by default. On a plain request, content carries the final answer directly with no separate reasoning trace.

Turn reasoning on by adding reasoning: true to the request, or use the :think suffix:

{ "model": "gemma/gemma-4-26b-a4b-it:think", "messages": [{ "role": "user", "content": "What is 17 times 23?" }], "max_tokens": 1024 }

// Or the reasoning field in the request body
{ "model": "gemma/gemma-4-26b-a4b-it", "reasoning": true, "messages": [{ "role": "user", "content": "What is 17 times 23?" }], "max_tokens": 1024 }

With reasoning on, the response carries a separate trace in message.reasoning_content; content still holds only the final answer. Reasoning tokens are counted in completion_tokens and billed as output, so keep max_tokens generous when reasoning is on.

Image input

The model accepts text and image together, and produces text only. Send the image as a base64 data URL:

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma/gemma-4-26b-a4b-it",
    "messages": [{
      "role": "user",
      "content": [
        { "type": "text", "text": "What is in this image?" },
        { "type": "image_url", "image_url": { "url": "data:image/png;base64,<base64>" } }
      ]
    }]
  }'

Audio and video input are not supported.

Tool calling

Tool calling is native. All three forms of tool_choice work — auto, required and a named function — and the returned tool_calls arguments are valid JSON. See Tool Calling for details.

curl "$LLMTR_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer llmtr-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma/gemma-4-26b-a4b-it",
    "messages": [{ "role": "user", "content": "What is the weather in Istanbul?" }],
    "tools": [{
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get current weather for a city",
        "parameters": {
          "type": "object",
          "properties": { "city": { "type": "string" } },
          "required": ["city"]
        }
      }
    }],
    "tool_choice": "auto"
  }'

Structured output

response_format accepts both json_object and json_schema:

{
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "capital",
      "schema": {
        "type": "object",
        "properties": {
          "country": { "type": "string" },
          "capital": { "type": "string" }
        },
        "required": ["country", "capital"]
      }
    }
  }
}

Data processing

This model is not hosted in Turkey. Requests are processed on third-party infrastructure abroad, and prompts and completions may be logged by the provider.

Limits

• Context window is 131,072 tokens (confirmed against live measurement).
• There is no separate output ceiling; the budget is shared. The prompt and the response share the same 131,072-token budget; no separate cap is applied to max_tokens beneath the context window. Measured at two different prompt sizes: the largest accepted max_tokens shrinks by exactly as much as the prompt grows.
• Do not size a request against usage.prompt_tokens — leave headroom. The provider decides whether to accept a request using its own count of the prompt, and that count can exceed the usage.prompt_tokens it later reports for the same request: in the 2026-08-24 measurement usage reported 4,023 tokens while the limit was computed against 5,011. A request whose max_tokens is set to reach exactly 131,072 based on the reported number can therefore be rejected unexpectedly; leave a few thousand tokens of headroom if you are targeting the total.
• An over-budget request returns 429, not 400. The rejection comes from the provider rather than the gateway, and its body reads exceeds its 131072-token context window. The same provider also returns 429 for transient slowness and for quota; those two carry retry after Ns and succeed when retried, whereas a size rejection does not.
• There is no prompt caching. Based on live measurements, no prefix prompt caching discount is offered.
• Audio and video input are not supported.