Models
Gemma 4 26B A4B
Google's fast model. Uses only a small part of itself per token, so it answers fast and cheaply.
This model isn't running right now. Requests to it get 404 with model_not_found. See Models for the ones you can use.
gemma-4-26b-a4b| Context window | 32,768 tokens (prompt and answer together) |
| Input | Text |
| Output | Text |
| Streaming | Yes |
| Prompt caching | Yes, automatic |
Tool calling (tools) | Yes |
Thinking (reasoning_effort) | Yes, off unless you ask for it |
| Price | Under Oversigt (Overview) on the platform |
| License | Apache 2.0 |
The model has 25 billion parameters but uses only 3.8 billion of them per token. That is why it answers faster and costs less than a model of the same size that uses all of them.
Use the model
curl https://api.danskllm.dk/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $DANSKLLM_API_KEY" \
-d '{
"model": "gemma-4-26b-a4b",
"messages": [
{"role": "system", "content": "Du er en hjælpsom assistent."},
{"role": "user", "content": "Forklar kort, hvad en API-nøgle er."}
]
}'Thinking
The model can think before it answers. That gives better answers on hard tasks, but the answer takes longer and costs more. Thinking can only be switched on or off, so every level gives the same thinking:
reasoning_effort | Means |
|---|---|
Not set, or "none" | No thinking. Fastest and cheapest. |
"low", "medium" or "high" | Thinking on. |
Other values get status 400 with param: "reasoning_effort".
curl https://api.danskllm.dk/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $DANSKLLM_API_KEY" \
-d '{
"model": "gemma-4-26b-a4b",
"reasoning_effort": "low",
"messages": [
{"role": "user", "content": "How many working days are there in November 2026?"}
]
}'What to know about thinking:
- It comes separately. The answer is in
contentas always. The thinking is in thereasoningfield next to it, and in a stream indelta.reasoning. The field is not part of OpenAI's API, so the SDKs don't have it in their types, but it is in the response. - It is paid as output. Thinking counts towards
completion_tokens. - It counts towards
max_tokens. If you setmax_tokenslow, thinking can use all of it, socontentis empty andfinish_reasonis"length". Setmax_tokenshigh, or leave it unset, when you use thinking.
Coding agents and other tools
Tools that can use an OpenAI-compatible provider, such as Cline, Continue and opencode, are set up like this:
| Setting | Value |
|---|---|
| Base URL | https://api.danskllm.dk/v1 |
| API key | Your key from the platform, dkllm-... |
| Model | gemma-4-26b-a4b |
| Context window | 32768 |
| Reasoning effort | none for quick steps, low when thinking should be on |
Set the context window in the tool to 32,768. Then the tool shortens the conversation before it gets too long, instead of getting 400 with context_length_exceeded.
Tool calls work as with OpenAI: send tools, and the model answers with tool_calls. Send the result back as a message with role: "tool".
By default an account can have 4 requests running at once. See Limits and fair use if your tool sends more at the same time.
More about parameters and streaming.