Models Capabilities / Gemma 4

gemma4-12b-direct

Gemma 4 12B with an adapter that declines fewer requests.

Updated · Live figures load from api.tokenflow.dev

Last 24 hours

Status
—
First token, median
—
Output
—
Success
—

Last 7 days

One row per UTC day; today so far in the last row.
DayFirst tokenOutputSuccess
————
————
————
————
————
————
————

Live figures cover all API traffic in the last 24 hours and refresh every minute. First token is the median time from request to the first streamed token, at any prompt length. Output is output tokens per second after the first token, over answers of 16 tokens or more. Success is the share of requests that did not fail on our side; client errors such as a bad request or an exhausted quota do not count against it. A figure needs at least 20 requests behind it; until then it shows a dash.

About

Gemma 4 12B, an open-weight model from Google DeepMind, served in NVFP4 with multi-token prediction on a DGX Spark. Smaller than Qwen3.8 27B, so first tokens and output arrive sooner; images cost about 280 prompt tokens each, whatever their size. This variant adds an adapter that makes the model decline fewer requests; otherwise it behaves as gemma4-12b.

Model IDgemma4-12b-direct
Context131,072 tokens, prompt plus output
InputText, Images
FeaturesTool calls
Streaming
Prompt caching
JSON mode and structured output
APIs/v1/chat/completions, /v1/responses, /v1/messages (Anthropic), /v1/completions
LimitsUp to 16 images per request, inline only (data URL or base64).
You are responsible for what you ask it to write.

Use it

Any OpenAI- or Anthropic-compatible client works: point it at https://api.tokenflow.dev with your API key and set the model to gemma4-12b-direct. For Claude Code and Codex, see coding agents; to try the model without code, open the console playground on it.

curl
curl https://api.tokenflow.dev/v1/chat/completions \
  -H "Authorization: Bearer $TOKENFLOW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gemma4-12b-direct", "messages": [{"role": "user", "content": "Hello"}]}'
Python, openai SDK
import os
from openai import OpenAI

client = OpenAI(base_url="https://api.tokenflow.dev/v1", api_key=os.environ["TOKENFLOW_API_KEY"])
reply = client.chat.completions.create(
    model="gemma4-12b-direct",
    messages=[{"role": "user", "content": "Hello"}],
)
print(reply.choices[0].message.content)