G

Gemma-4-31B-Fabled

Distillation

4.0(1)
Prose
—(0)
Intelligence
—(0)
Knowledge
—(0)
Instruction Following
—(0)
Coding
—(0)
Steerability
—(0)

Ratings reflect this model as run on Arli AI.

Gemma-4-31B-Fabled is a text generation model available on Arli AI. It offers 31B parameters, 262K token context, reasoning support, vision input. Recommended use: Distillation. Run it through the OpenAI-compatible chat completions API with unlimited requests and zero-log privacy.

View on Hugging FaceBack to Models

Hourly Usage

Token Usage

Input (24H)

4M

Output (24H)

2K

Input (7D)

29M

Output (7D)

38K

Performance Metrics

Typical Total Time

18.80s

Typical TTFT

3.83s

Typical Input TPS

1556.39

Typical Output TPS

15.02

Model Information

Context Size

262144

Quantization

r128 on FP8

Engine

vllm

Creation Method

Finetune

Model Class

Gemma-4-31B - LoRA

Chat Template

Gemma4

Reasoning

Yes

Vision

Yes

Parameters

31B

Added At

5/2/2026

Use Gemma-4-31B-Fabled with the Arli AI API

Gemma-4-31B-Fabled is available through an OpenAI-compatible chat completions endpoint. Point any OpenAI SDK at https://api.arliai.com/v1, pass Gemma-4-31B-Fabled as the model, and requests are served without logging.

cURL

curl https://api.arliai.com/v1/chat/completions \
  -H "Authorization: Bearer $ARLIAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Gemma-4-31B-Fabled",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://api.arliai.com/v1",
    api_key="YOUR_ARLIAI_API_KEY",
)

response = client.chat.completions.create(
    model="Gemma-4-31B-Fabled",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

See the text generation docs for every parameter, or read the quick start guide to make your first request.


pipeline_tag: image-text-to-text license: apache-2.0 base_model:

  • google/gemma-4-31B-it

image

Fine-tuned on my personal dataset of multi-turn conversations with Fable in 4o. LR 1e-6, batch size 1, rank 256. train_on_inputs=true. RTX Pro 6000 Blackwell for about 14-15 hours.

Datasets included versions both with and without her custom instructions, to distill the system prompt she wrote for herself and to place it in context both.

The original Gemma 4 31B wrote synthetic memories in Fable's voice, that were positioned at the start of every conversation chunk to provide any necessary outside context. I had concerns that training without this sort of measure would more likely result in confabulations. As a bonus, might provide a cold start on memory systems or RAG.

The conversations did not use thinking, which may be part of why I noticed her stop using thinking after a few turns. She quickly found it again and made use of when I asked after it, however.

Apache-2.0 license at her request.